Running Untrusted Application Code: Sandboxing

Running untrusted code

! We often need to run buggy/unstrusted code:
n

programs from untrusted Internet sites:

w toolbars, viewers, codecs for media player

old or insecure applications:

ghostview, outlook

legacy daemons: sendmail, bind honeypots if application misbehaves, kill it

! Goal:

Approach: confinement
! Confinement: ensure application does not deviate from
pre-approved behavior

! Can be implemented at many levels:

n

Hardware: run application on isolated hw (air gap) w difficult to manage Virtual machines: isolate OS s on single hardware System call interposition: w Isolates a process in a single operating system Isolating threads sharing same address space: w Software Fault Isolation (SFI) Application specific: e.g. browser-based confinement

n n

Implementing confinement
! Key component:
n

reference monitor

Mediates requests from applications w Implements protection policy w Enforces isolation and confinement Must always be invoked: w Every application request must be mediated Tamperproof: w Reference monitor cannot be killed w or if killed, then monitored process is killed too Small enough to be analyzed and validated

A simple example:
! To use do: (must be root)
chroot /tmp/guest su guest

chroot

! Often used for guest accounts on ftp sites

root dir / is now /tmp/guest EUID set to guest

! Now

/tmp/guest applications in jail

is added to file system accesses for r )

open( /etc/passwd ,

r )

open( /tmp/guest/etc/passwd , application cannot access files outside of jail

Jailkit
Problem: all utility progs (ls, ps, vi) must live inside jail jailkit project: auto builds files, libs, and dirs needed in jail environment jk_init: creates jail environment jk_check: checks jail env for security problems checks for any modified programs, checks for world writable directories, etc. jk_lsh: restricted shell to be used inside jail note: simple chroot jail does not limit network access

Escaping from jails

! Early escapes:
relative paths r ) r ) open( ../../etc/passwd ,

open( /tmp/guest/../../etc/passwd ,

! chroot should only be executable by root

n

otherwise jailed app can do: w create dummy file /aaa/etc/passwd w run chroot /aaa w run su root to become root (bug in Ultrix 4.0)

Many ways to escape jail as root

! Create device that lets you access raw disk ! Send signals to non chrooted process ! Reboot system ! Bind to privileged ports

Freebsd jail
! Stronger mechanism than simple chroot ! To run:
jail jail-path hostname IP-addr cmd
n n

calls hardened chroot

(no

../../

escape)

can only bind to sockets with specified IP address and authorized ports can only communicate with process inside jail root is limited, e.g. cannot load kernel modules

n n

Problems with chroot and jail

! Coarse policies:
n n

All or nothing access to file system Inappropriate for apps like web browser w Needs read access to files outside jail (e.g. for sending attachments in gmail)

! Do not prevent malicious apps from:

n n

Accessing network and messing with other machines Trying to crash host OS

System call interposition: a better approach to confinement

Sys call interposition

! Observation: to damage host system (i.e. make
persistent changes) app must make system calls n To delete/overwrite files: unlink, open, write n To do network attacks: socket, bind, connect, send

! Idea:
n

monitor app system calls and block unauthorized calls

! Implementation options:
n n n

Completely kernel space (e.g. GSWTK) Completely user space (e.g. program shepherding) Hybrid (e.g. Systrace)

Initial implementation
! Linux ptrace:

(Janus)

process tracing tracing process calls: ptrace ( , pid_t pid , ) and wakes up when pid makes sys call.
user space

monitored application (outlook) open( etc/passwd , r )

monitor

OS Kernel

! Monitor kills application if request is disallowed

Complications
! If app forks, monitor must also fork
n

Forked monitor monitors forked app

! If monitor crashes, app must be killed ! Monitor must maintain all OS state associated with app
n n

current-working-dir (CWD), UID, EUID, GID Whenever app does cd path monitor must also update its CWD w otherwise: relative path requests interpreted incorrectly

Problems with ptrace

! Ptrace too coarse for this application
n

Trace all system calls or none w e.g. no need to trace close system call Monitor cannot abort sys-call without killing app

! Security problems: race conditions

n

Example: symlink: me -> mydata.dat proc 1: open( me ) monitor checks and authorizes not atomic proc 2: me -> /etc/passwd OS executes open( me ) Classic TOCTOU bug: time-of-check / time-of-use

time
n

Alternate design: systrace

user space monitored application (outlook) open( etc/passwd , sys-call gateway r ) monitor policy file for app

systrace permit/deny

OS Kernel

! systrace only forwards monitored sys-calls to monitor

(saves context switches) ! systrace resolves sym-links and replaces sys-call path arguments by full path to target ! When app calls execve, monitor loads new policy file

Policy
! Sample policy file:
path allow /tmp/* path deny /etc/passwd network deny all

! Specifying policy for an app is quite difficult

n

Systrace can auto-gen policy by learning how app behaves on good inputs If policy does not cover a specific sys-call, ask user but user has no way to decide

! Difficulty with choosing policy for specific apps (e.g.

browser) is main reason this approach is not widely used

NaCl: a modern day example

Browser HTML JavaScript NaCl runtime Quake NPAPI

! Quake: untrusted x86 code ! Two sandboxes:

n n

outer sandbox: restricts capabilities using sys call interposition Inner sandbox: uses x86 memory segmentation to isolate application memory from one another

Confinement using Virtual Machines

Virtual Machines
VM2
Apps
Apps
VM1

Guest OS 2
Host OS
Hardware
Example: NSA NetTop

Guest OS 1

Virtual Machine Monitor (VMM)

single HW platform used for both classified and unclassified data

Why so popular now?

! VMs in the 1960 s:
n n

Few computers, lots of users VMs allow many users to shares a single computer non-existent

! VMs 1970 s 2000: ! VMs since 2000:

n

Too many computers, too few users w Print server, Mail server, Web server, File server, Database server, Wasteful to run each service on a different computer w VMs save hardware while isolating services More generally: VMs heavily used in cloud computing

VMM security assumption

! VMM Security assumption:
n n

Malware can infect guest OS and guest apps But malware cannot escape from the infected VM

w Cannot infect host OS

w Cannot infect other VMs on the same hardware

! Requires that VMM protect itself and is not buggy

n n

VMM is much simpler than full OS but device drivers run in Host OS

Problem: covert channels

! Covert channel:
unintended communication channel between isolated components n Can be used to leak classified data from secure component to public component Classified VM
secret doc

Public VM covert channel listener

malware

VMM

An example covert channel

! Both VMs use the same underlying hardware ! To send a bit b {0,1} malware does:
n n

b= 1: at 1:30.00am do CPU intensive calculation b= 0: at 1:30.00am do nothing

! At 1:30.00am listener does a CPU intensive calculation

and measures completion time n Now b=1 completion-time > threshold

! Many covert channel exist in running system:

n n

File lock status, cache contents, Very difficult to eliminate

interrupts,

VMM Introspection:

[GR 03]

protecting the anti-virus system

Intrusion Detection / Anti-virus

! Runs as part of OS kernel and user space process
n n

Kernel root kit can shutdown protection system Common practice for modern malware run IDS system in the network Problem: insufficient visibility into user s machine
(protected from malware)

! Standard solution:
n

! Better: run IDS as part of VMM

n n

VMM can monitor virtual hardware for anomalies VMI: Virtual Machine Introspection
w Allows VMM to check Guest OS internals

Sample checks
Stealth malware: n Creates processes that are invisible to ps n Opens sockets that are invisible to netstat 1. Lie detector check n Goal: detect stealth malware that hides processes and network activity n Method:
w VMM lists processes running in GuestOS w VMM requests GuestOS to list processes (e.g. ps) w If mismatch, kill VM

Sample checks
2. Application code integrity detector n VMM computes hash of user app-code running in VM n Compare to whitelist of hashes w Kills VM if unknown program appears 3. Ensure GuestOS kernel integrity n example: detect changes to sys_call_table 4. Virus signature detector n Run virus signature detector on GuestOS memory 5. Detect if GuestOS puts NIC in promiscuous mode

Subvirt: subvirting VMM confinement

Subvirt
! Virus idea:
n n n

Once on the victim machine, install a malicious VMM Virus hides in VMM Invisible to virus detector running inside VM Anti-virus

HW

Anti-virus

OS HW

OS
VMM and virus

The MATRIX

VM Based Malware (blue pill virus)

! VMBR:
a virus that installs a malicious VMM (hypervisor)
(Oct, 2006) http://www.microsoft.com/whdc/system/platform/virtual/ CPUVirtExt.mspx
n

! Microsoft Security Bulletin:

Suggests disabling hardware virtualization features by default for client-side systems

! But VMBRs are easy to defeat

n

A guest OS can detect that it is running on top of VMM

VMM Detection
! Can an OS detect it is running on top of a VMM? ! Applications:
n n

Virus detector can detect VMBR Normal virus (non-VMBR) can detect VMM w refuse to run to avoid reverse engineering Software that binds to hardware refuse to run on top of VMM
(e.g. MS Windows)

can

DRM systems may refuse to run on top of VMM

VMM detection
n

(red pill techniques)

1. VM platforms often emulate simple hardware

VMWare emulates an ancient i440bx chipset but report 8GB RAM, dual Opteron CPUs, etc.

2. VMM introduces time latency variances n Memory cache behavior differs in presence of VMM n Results in relative latency in time variations for any two operations 3. VMM shares the TLB with GuestOS n GuestOS can detect reduced TLB size and many more methods
[GAWF 07]

VMM Detection
Bottom line: The perfect VMM does not exist

! VMMs today (e.g. VMWare) focus on:

Compatibility: ensure off the shelf software works Performance:

minimize virtualization overhead

! VMMs do not provide transparency

n

Anomalies reveal existence of VMM

Software Fault Isolation

Software Fault Isolation

! Goal:
n n

confine apps running in same address space Codec code should not interfere with media player Device drivers should not corrupt kernel

! Simple solution: runs apps in separate address spaces

n

Problem: slow if apps communicate frequently w requires context switch per message

Software Fault Isolation

! SFI approach:
n

Partition process memory into segments

data segment code segment data segment

code segment

app #1

app #2

Locate unsafe instructions: jmp, load, store w At compile time, add guards before unsafe instructions w When loading code, ensure all guard are present

Segment matching technique

! Designed for MIPS processor. Many registers available. ! dr1, dr2: dedicated registers not used by binary
n n

Compiler pretends these registers don t exist not Guard ensures code does dr2 contains segment ID R12 [addr]

load data from another segment

! Indirect load instruction

becomes:

dr1 addr scratch-reg (dr1 >> 20) compare scratch-reg and dr2 trap if not equal R12 [addr]

: get segment ID : validate seg. ID : do load

Address sandboxing technique

! dr2:
holds segment ID R12 [addr] : zero out seg bits : set valid seg ID : do load

! Indirect load instruction

becomes:

dr1 addr & segment-mask dr1 dr1 | dr2 R12 [dr1]

! Fewer instructions than segment matching

but does not catch offending instructions

! Lots of room for optimizations: reduce # of guards

Cross domain calls

caller domain
call draw stub

callee domain
draw: return

stub

br addr br addr br addr

! Only stubs allowed to make cross-domain jumps ! Jump table contains allowed exit points from callee
n

Addresses are hard coded, read-only segment

SFI: concluding remarks

! For shared memory: use virtual memory hardware
n

Map same physical page to two segments in addr space

! Performance
n

Usually good:

mpeg_play, 4% slowdown

! Limitations of SFI: harder to implement on x86 :

n n n

variable length instructions: unclear where to put guards few registers: can t dedicate three to SFI many instructions affect memory: more guards needed

Summary
! Many sandboxing techniques:
n n n n n

Physical air gap, Virtual air gap (VMMs), System call interposition Software Fault isolation Application specific (e.g. Javascript in browser)

! Often complete isolation is inappropriate

n

Apps need to communicate through regulated interfaces

! Hardest aspect of sandboxing:

n

Specifying policy:

what can apps do and not do

THE END