Cookies Lack Integrity: Real-World Implications

Xiaofeng Zheng, Tsinghua University; Jian Jiang, University of California, Berkeley;

Jinjin Liang, Tsinghua University; Haixin Duan, Tsinghua University and International
Computer Science Institute; Shuo Chen, Microsoft Research Redmond;
Tao Wan, Huawei Canada; Nicholas Weaver, International Computer Science Institute
and University of California, Berkeley
https://www.usenix.org/conference/usenixsecurity15/technical-sessions/presentation/zheng

This paper is included in the Proceedings of the

24th USENIX Security Symposium
August 1214, 2015 Washington, D.C.
ISBN 978-1-931971-232

Open access to the Proceedings of

the 24th USENIX Security Symposium
is sponsored by USENIX

Cookies Lack Integrity: Real-World Implications

Xiaofeng Zheng1,2,3 , Jian Jiang7 , Jinjin Liang1,2,3 , Haixin Duan1,3,4 , Shuo Chen5 , Tao Wan6 , and
Nicholas Weaver4,7
1 Institute

for Network Science and Cyberspace, Tsinghua University

of Computer Science and Technology, Tsinghua University
3 Tsinghua National Laboratory for Information Science and Technology
4 International Computer Science Institute
5 Microsoft Research Redmond
6 Huawei Canada
7 UC Berkeley

2 Department

Abstract
A cookie can contain a secure flag, indicating that it
should be only sent over an HTTPS connection. Yet there
is no corresponding flag to indicate how a cookie was
set: attackers who act as a man-in-the-midddle even temporarily on an HTTP session can inject cookies which
will be attached to subsequent HTTPS connections. Similar attacks can also be launched by a web attacker from a
related domain. Although an acknowledged threat, it has
not yet been studied thoroughly. This paper aims to fill
this gap with an in-depth empirical assessment of cookie
injection attacks. We find that cookie-related vulnerabilities are present in important sites (such as Google and
Bank of America), and can be made worse by the implementation weaknesses we discovered in major web
browsers (such as Chrome, Firefox, and Safari). Our
successful attacks have included privacy violation, online victimization, and even financial loss and account
hijacking. We also discuss mitigation strategies such as
HSTS, possible browser changes, and present a proof-ofconcept browser extension to provide better cookie isolation between HTTP and HTTPS, and between related
domains.

Introduction

The same-origin policy (SOP) is a corner stone of web

security, guarding the web content of one domain from
the access from another domain. The most standard definition of origin is a 3-tuple, consisting of the scheme,
the domain and the port number. However, the notion of
origin regarding cookies is fairly unusual cookies are
not separated between different schemes like HTTP and
HTTPS, as well as port. The domain isolation of cookie
is also weak: different but related domains can have a
shared cookie scope. A cookie may have a secure flag,
indicating that it should only be presented over HTTPS,
ensuring confidentiality of its value against a network

USENIX Association

man-in-the-middle (MITM). However, there is no similar

measure to protect its integrity from the same adversary:
an HTTP response is allowed to set a secure cookie for
its domain. An adversary controlling a related domain
is also capable to disrupt a cookies integrity by making
use of the shared cookie scope. Even worse, there is an
asymmetry between cookies read and write operations
involving pathing, enabling more subtle form of cookie
integrity violation.
The lack of cookie integrity is a known problem,
noted in the current specification [2]. However, the
real-world implications are under-appreciated. Although
the problem has been discussed by several previous researchers [4, 5, 30, 32, 24, 23], none provided in-depth
and real-world empirical assessment. Attacks enabled by
merely injecting malicious cookies could be elusive, and
the consequence could be serious. For example, a cautious user might only visit news websites at open wireless
networks like those at Starbucks. She might not know
that this is sufficient for a temporary MITM attacker to
inject malicious cookies to poison her browser, and compromise her bank account when she later logs on to her
bank site at home.
We aim to understand how could attackers launch
cookie inject attacks, and what are the damaging consequences to real-world websites. Our study shows
that most websites are potentially susceptible to cookie
injection attacks by network attackers. For example,
only one site in the Alexa top 100 websites has fully
deployed HTTP Strict Transport Security (HSTS) on
its top-level domain, a sufficient server-side protection
to counter cookie injection attacks by network attackers (Section 3). We also found a number of browser
vulnerabilities and implementation quirks that can be exploited by cookie injection attacks (Section 4). Notably,
all major browsers, except Internet Explorer (IE), respect
the Set-Cookie header in a 407-response (i.e., an Authentication Required Response) when configured to use
a proxy. Because of this vulnerability, even websites

24th USENIX Security Symposium 707

adopting sufficient HSTS are subject to cookie injection

attacks by a malicious proxy.
Our study also shows that current cookie practices
have widespread problems when facing cookie injection
attacks (Section 5). We demonstrate multiple exploitations against large websites. For example, we show that
an attacker can put his Gmail chat gadget on a victims
screen without affecting the victims use of Gmail and
other Google services. We also demonstrate that an attacker can hijack a victims online deposit to his account,
or even deliver the victims online purchase to his address. Other exploitations include user tracking, crosssite scripting (XSS) attacks against large financial sites
embedded in injected cookies, etc..
We have developed a mitigation strategy (Section 6).
By modifying how browsers treat secure cookies, it is
possible to largely mitigate cookie injection attacks by
network attackers. We have also considered possible
browser enhancements to mitigate cookie injection from
web attackers. We implement our proposals as a proofof-concept browser extension. A preliminary evaluation
does not encounter compatibility issues.
In summary, this work makes the following main contributions:

2.1

Background
Cookies

Cookies are a browser-side assisted state management

mechanism that are pervasively used by web applications [2]. Cookies can be set by either HTTP servers
using Set-Cookie: header or client side JavaScript
with a write to document.cookie. A cookie can
have five optional attributes: domain and path specifying the cookies scope; expires stating when it should
be discarded; secure specifying that it should only be
sent over HTTPS connections, and HTTPOnly preventing browser-side scripts from reading the cookie. When
sending a request to a server, a web browser includes all
unexpired cookies whose domains and paths match the
requested URL, excluding those marked as secure from
the inclusion in an HTTP request.
Cookies have two fairly unusual behaviors. First,
there is a critical disconnection between cookie storage and reading. Cookies are set and stored as a
name/domain/path to value attributes mapping, but only
name-value pairs are presented to both JavaScript and
web servers. This asymmetry allows cookies with the
same name but different domain and/or path scopes to be
written into browser; a subsequent reader can read out
all same name cookies together, yet cannot distinguish
them because the other attributes such as path are not
presented in the reading process. Another complication
occurs when writing a cookie, the writer can specify arbitrary value for the path attribute, not limited by the
URL of the writers context.
Moreover, the security policy for cookies is not as
stringent as the classic SOP. In web security, the SOP
is the most important access control mechanism to segregate static contents and active scripts from different
origins [3]. An origin for a given URL is defined by a
3-tuple: scheme (or protocol), e.g. HTTP or HTTPS, domain (or host), and port (not supported by IE (Internet
Explorer)). However, the security policy guarding cookies does not provide separation based on either scheme or
port but only on domain [2]. In addition, a website can
set cookies with flexible domain scopes: 1) not shared
(i.e., host-only), 2) shared with its subdomains, or 3)
shared with its sibling domains (i.e., using its parent domain as the scope). For the third case, a restriction is enforced by browser to ensure that a cookie domain scope
is not too wide. For example, www.example.com can
set a cookie with the scope of .example.com, but it cannot set a cookie with .com as the scope because .com is a
public top level domain (TLD). Unfortunately, there is no
clear definition of whether a domain scope is too wide
(See Section 3.2).
The combination of the read/write asymmetry and the

We provided an evaluation of potential susceptible

websites to cookie injection attacks, including a detailed measurement of full HSTS adoption and an
assessment of shared domains used by Content Delivery Networks (CDNs).
We examined both browser-side and server-side
cookie implementation, in which we found several browser vulnerabilities and a number of nonconforming and/or inconsistent implementations
that could be exploited in cookie injection attacks.
We demonstrated the severity and prevalence of
cookie injection attacks in the real world. In particular, our exploitations against a variety of large
websites show that cookie injection enables complicated interactions among implements, applications,
and various known attacks.
We developed and implemented browser-side enhancements to provide better cookie isolation. Our
evaluation showed promising results in compatibility.
Together, this work provides a close-up picture of the
cookie integrity problem and the threats of cookie inject
attacks. We intend to provide a context for motivating
further discussion in research community and industry.
2
708 24th USENIX Security Symposium

USENIX Association

lack of domain or scheme segregation implies that a domain cannot protect the integrity of its cookie from an active MITM or a malicious/compromised related domain
that shares some cookie domain scope with it. There are
two forms of cookie integrity violations:

in its HSTS header, a browser will allow HTTP connection to foo.example.com. Worse, even if the HSTS
policy of example.com specifies includeSubDomains,
this will not be checked by a browser if a user only visits
bar.example.com unless the page includes a reference
to example.com.

Cookie Overwriting. If a cookie shares the domain

scope with a related domain, it can be directly overwritten by that domain using another cookie with
the exactly same name/domain/path. Of particular
note, although a secure cookie can only be read by
an HTTPS process, it can be written or overwritten
by an HTTP request.

2.3

It is a known vulnerability that cookies can be injected

by HTTP response into subsequent HTTPS request, and
from one domain to another related domain. Johnston
and Moore reported such problem in 2004 [19]. Their
report already pinpointed the root cause: the loosely
defined SOP for cookies. Unfortunately browsers vendors did not fix the problem probably because they were
concerned of potential incompatibility issues. In 2008,
Evans described an attack called cookie forcing that exploits cookie integrity deficiency to overwrite cookies in
HTTPS sessions [7]. In 2013, GitHub migrated their domain for hosting users homepages from github.com to
github.io after they recognized the threat of cookie injection from/to a shared domain whose subdomains belong to mutually untrusted users; they described detailed
steps of several possible cookie injection exploits and referred to them as cookie tossing [11].
The problem was also noted in several more formal
publications. Barth et al. discussed security implications of cookie overwriting on session initiation [4].
They also proposed a new header Cookie-Integrity
to provide additional information so that web server can
distinguish between cookies set from HTTP and those
set from HTTPS. Bortz et al. also reviewed the problem
and proposed a new header Origin-Cookie that guarantees integrity by enforcing a complete 3-tuple SOP [5].
Singh et al. referred the difference between the classic
SOP and the cookie SOP as inconsistent principal labeling [30]. Both Zalewskis book [32] and the current
cookie specification by Barth [2] explained the cookie
integrity deficiencies in great detail. We also learned of
two technical reports, one from Black Hat EU by Lundeen [23] and the other from Black Hat AD by Lundeen et al. [24], that illustrated several subtle attacks initiated by cookie injection.
Although a known threat, previous research fall short
of in-depth empirical assessment of its real-world security implications. This work aims to fill this gap. We
provide a detailed comparison in Section 7.

Cookie Shadowing. Alternatively, an attacker with

the control of a related domain can intentionally
shadow a cookie by injecting another one that has
the same name, but different domain/path scope.
For example, to shadow a cookie with value=good;
domain=www.example.com; path=/; secure, a
related domain evil.example.com can write a
cookie with value=bad; domain=.example.com;
path=/home. Later, when browser issues a request to https://www.example.com/home, both
cookies match the URL and are included. For
most browsers, the cookie header will be Cookie:
value=bad; value=good;. The good cookie could
be shadowed by the bad one if a website happens
to prefer the value of bad over good.
Note while the good cookie has a secure flag and
is sent over HTTPS, it can still be shadowed with a
cookie set from an HTTP connection.

2.2

Cookie Injection Attacks

HSTS

HSTS (HTTP Strict Transport Security) allows a server

to inform a client to only initiate communications over
HTTPS. It was originally proposed by Jackson and
Barth to address a number of MITM threats such as
cookie sniffing and SSL stripping [18], and is now
standardized in RFC6797 as a HTTP response header
Strict-Transport-Security [15].
The HSTS header requires a max-age attribute
indicating how long a browser should keep the
HSTS policy for that domain. An optional attribute
includeSubDomains tells a browser to apply the HSTS
policy to its all subdomains. After receiving an HSTS
header, a conforming browser ensures that all subsequent connections to that domain always take place over
HTTPS until the policy expires. Chrome and Firefox also
support a preloaded list that contains self-declared websites supporting HSTS. For more information on HSTS,
please see [22].
HSTS coverage can often be incomplete. For example,
if example.com does not specify includeSubDomains

Threat Analysis

We first present the threat model for cookie injection attacks. For each type of attacker, we analyze its real-world
threat. Table 1 gives an overview.
3

USENIX Association

24th USENIX Security Symposium 709

Attacker
Network
Attacker

Active MITM
Malicious Proxy

Web
Attacker

Full control of related domain

XSS on related domain

Root Cause

Attack Surface

SOP without protocol &

complete domain isolation.

Websites and browsers that allow attackers to reply an unencrypted request to a related domain with forged response.

Mitigation

SOP without complete domain isolation.

Websites using shared domains.

Websites with compromised related domains.

Full HSTS
Public suffix list
Out of scope

Table 1: Overview of the threats of cookie injection attacks

3.1

Threat Model

Domain Ranking
<10 10-102 102 -103 103 -104 104 -105 > 105
Valid HTTPS
7
52
353
2,914 20,548 128,805
Full HSTS
0
1
7
35
212
997

Two classes of attackers can manipulate a target sites

cookies: an active network adversary or a remote adversary able to host or inject content on a related domain.
The active MITM attacker (including the classic
MITM fully controlling the network and the Man-on-theSide (i.e., wiretapping and packet-injecting)) can load arbitrary cookies through HTTP into the targets cookie
store. The attacker modifies an unrelated HTTP request to create a hidden iframe in a web page. The attackers iframe then creates a series of HTTP fetches to
the target domains, which the attacker responds to with
Set-Cookie headers to poison the victims cookie store.
A malicious proxy is at least as powerful as an active
MITM in terms of manipulating network traffic. Moreover, because the browser has extra protocol interactions
with the proxy, potential logic flaws or implementation
bugs might give the malicious proxy additional chances
to break in. Chen et al. highlighted this threat with a
number of logic flaws [6]. Our study also targets this
type of issues related to unexpected capabilities for a malicious proxy to inject cookies.
Finally, if an attacker controls a related domain directly, he may launch cookie injection remotely. The attacker does not need full control of the related server, just
the ability to host JavaScript. This attacker cannot target
arbitrary domains, but can target any other domain under
the same top level domain.
One key property of all these adversaries is its ability
to change state. For example, a victim might only visit
her bank from known-good networks, but an attacker can
poison the victims browser when the victim is on an
open wireless network. Only later, when the victim has
now returned to the safe network and visits her bank,
does the attack actually affect the victim.

3.2

Table 2: Ranking distribution of domains with valid

HTTPS and full HSTS.
considered non-public) with the includeSubDomains
option, which we refer to as full HSTS; 2) the browser
supports HSTS; and 3) the browser has received the full
HSTS policy from the base domain of the target domain.
Unfortunately, the support and adoption of HSTS in
the real world is unsatisfactory. First, all current versions of IE, a major browser with considerable marketshare, do not support HSTS (Microsoft announced
that its new browser will support HSTS [16]). Second, there is limited adoption of full HSTS among sites.
We scanned 961,857 base domains from the Alexa top
one million websites and also examined if these domains present in the Chromes preloaded HSTS list [28].
While we observed 152,679 (15.87%) domains have deployed HTTPS with valid certificates, we only found
1,252 (0.13%) domains have enabled full HSTS. Moreover, most of the full HSTS domains are low ranked domains (see Table 2). A recent study by Kranch and Bonneau also presented a similar total number of full HSTS
domains among the Alexa top one million websites [22].
Because of the prevalence of unsafe networks like
open wireless networks and the very limited deployment/availability of full HSTS protection, we consider
cookie injection by active network attackers a pervasive
and severe threat, especially for websites who have deployed HTTPS to prevent active network attackers from
launching other possible attacks such as eavesdropping
or active script injection, yet have not enabled full HSTS.
Web Attack Surface: Generally, a web attacker
might be able to control a related domain in two ways.
First, for large websites that all subdomains are used internally, an attacker can fully control one subdomain by
compromising its DNS resolution or its hosting server.
The attacker can also exploit a XSS vulnerability on a
subdomain of a large website. A cookie injection attack
can then be launched to target other subdomains.
A greater concern is when a website either hosts user
content or shares a domain scope with other possibly
untrustworthy sites. This problem is inherent from the

Attack Surface

Network Attack Surface: The only current protection

against an active network attacker requires that the victims browser never issues an unencrypted HTTP connection to a target site or any related domain. This condition holds if 1) the target domain enables HSTS on its
base domain 1 (i.e. the first upper-level domain that is
1 We

learned this term from Kranch and Bonneaus recent HSTS

study [22].

4
710 24th USENIX Security Symposium

USENIX Association

Vendor

Domain
Publis Suffix List? Vulnerable?
akamai.net
No
n/a 1
akamaiedge.net
No
n/a
Akamai
akamaihd.net
No
n/a
edgesuite.net
No
n/a
msecnd.net
No
Yes
Azure
windows.net
No
Yes
BitGravity
bitgravity.com
No
n/a
CacheFly 2
cachefly.net
No
Yes
CDN77
cdn77.net
No
Yes
CDNetworks
cdngc.net
No
n/a
CDN.net
worldcdn.net
No
n/a
ChinaCache
chinacache.net
No
n/a
ChinaNetCenter wscloudcdn.com
No
n/a
CloudFlare 3
cloudflare.net
No
Yes
CloudFront
cloudfront.net
Yes
No
EdgeCast
edgecasecdn.net
No
n/a
Exceda
expresscdn.com
No
Yes
Fastly 3
fastly.net
Yes
Yes
Highwinds
hwcdn.net
No
n/a
Incapsula
incapsula.net
No
Yes
Internap
internapcdn.net
No
n/a
Jiasule
jiashule.com
No
Yes
KeyCDN 2
kxcdn.com
No
Yes
Level3
footprint.net
No
n/a
Limelight
linwd.net
No
n/a
MaxCDN
netdna-cdn.com
No
Yes
Squixa
squixa.net
No
n/a

weaker cookie SOP. As we previously discussed in Section 2.1, a domain is allowed to set cookies with wider
domain scope as long as the scope is not considered
public. Hence, a clear boundary between public and
non-public domain scope is needed to prevent cookie
injection from undesired shared cookie domain. However, this is not easy to define and implement clearly.
First, many top-level domains (a.k.a., TLDs), especially
country code top-level domains (a.k.a., ccTLDs) have
their own reserved suffixes such as .com.cn, .co.uk,
which are mostly TLD-specific. Second, many websites
use shared domains to assign subdomains to their mutually untrusted clients. Such shared domain providers
include cloud hosting providers, web hosting providers,
blog providers, CDN providers etc. These shared domains should also be considered as non-public in terms
of cookie domain scope.
The problem of cookie domain scope boundary is
partially remedied by a community effort initiated by
Mozilla called public suffix list, which maintains an
exceptional list containing TLDs, TLD-reserved suffixes,
and self-declared shared domains [25]. Public information suggests that the list is enforced by major browser
vendors including IE, Chrome, Firefox, and Opera, while
our own tests confirm that Safari implements this list.
Our study of the public suffix list shows that the
public shared domains list still exposes an attack surface for cookie injection. First, we empirically identified 45 shared domains from the Alexa top one million
websites, among which only 10 Google domains and 3
non-Google domains are included in the public suffix
list. Among the remaining domains, we found at least
4 domains (sinaapp.com, weebly.com, myshopify.
com, and forumotion.com) allow customized serverside code or browser-side scripts. Websites hosting on
these domains are vulnerable to cookie injection attacks.
Another easy-to-miss corner case is shared domains
used by CDNs. CDNs commonly assign subdomains
or sub-directories of shared domains to their customers.
If a website directly uses a shared domain assigned by
its CDN provider, and the CDN provider does not handle the shared domain carefully, then the website is subject to cookie injection attacks from malicious customers
of the same CDN provider. While websites rarely use
shared domains as their main domains, a common practice is to refer static resources (e.g., JavaScript files, images) using shared domain URLs. Although cookies under these resource URLs are usually not processed by
server-side code or browser-side scripts, cookie injection
attacks could still cause serious consequences. For example, suppose both websites A and B host their static
resource files under one shared domain from the same
CDN. Website A can inject garbage cookies from the
requests to his resource files with specific paths so that

1: n/a refers to the case that we were not able to test.

2: CDNs attempting to defend cookie related attacks on shared domains by filtering the Set-Cookie header.
3: CDNs allowing shared cookie scopes in customer-specific prefixes
of shared domains.

Table 3: Assessment of cookie injection attacks on

shared domains used by CDNs.
the injected cookies will be sent with the requests of resource files to website B. This type of cookie injection
attack could cause performance downgrade, bandwidth
consumption, and even denial-of-service (DoS) if the
amount of injected cookies exceeds the servers header
size limitation2 . In worst case, DoS of a critical resource
file like a JavaScript library could break the whole website.
We empirically collected 28 shared domains used
by 23 CDNs 3 , in which only 2 domains are registered in the public suffix list, as presented in Table 3. We were also able to sign up and test 13
shared domains from 12 CDNs. While we confirmed
that cloudfront.net is immune because of its presence in the public suffix list, for each of the other 12
domains, we successfully launched DoS attack on one
test URL by injecting 72KB cookies from another test
URL. Our experiments also found two problematic behaviors. First, CacheFly and KeyCDN attempt to defend cookie related attacks by filtering the Set-Cookie
header in response instead of utilizing the public suf2 Although the current HTTP specification does not define any limitation on the size of request header [9], most of web server implementations do so by default. For example, nginx by default limits a single
HTTP header not to exceed 8KB [26].
3 We collected most of the CDNs from http://www.cdnplanet.
com/cdns/.

5
USENIX Association

24th USENIX Security Symposium 711

fix list, which fails to prevent JavaScript from injecting cookies. Second, although Fastly has declared several subdomains of fastly.net as public suffix, its
naming mechanism enables shared scopes in customerspecific prefixes, making its customers still vulnerable to
cookie injection attacks. For example, for a customer
foo.com, Fastly assigns a customer subdomain foo.
com.global.prod.fastly.net. Although the suffix
global.prod.fastly.net is present in the public suffix list, the prefix causes a cookie scope com.global.
prod.fastly.net shared with other customer subdomains such as bar.com.global.prod.fastly.net.
CloudFlare also has the same problem. We have reported
this problem to all vulnerable vendors. CloudFlare and
CDN77 have acknowledged our reports. The response
from CloudFlare said that they are considering to disable
direct access of all cloudflare.net URLs to defend
against this problem.

with the same name, while the name-value map interface in Python standard library prefers the last-ordered
cookie, all others prefers the first-ordered one. This explains why cookie shadowing is possible and the example
given in Section 2.1 works in many cases.
Cookie Storage Limitation. The specification has
several vague suggestions for browsers to limit the number and size of stored cookies. We found all major
browsers set the maximum size of a single cookie to 4
KB. Chrome, Firefox, and Opera implements a cookie jar
for every base domain, with the total numbers of cookies
limited to 180, 150, and 180, respectively. IEs cookie jar
implementation is per cookie domain scope, with the total number of cookies limited to 50. We did not reach Safaris cookie storage limit after writing and reading 1,000
cookies.
Cookie Header Size Limit. While Safari does not
seem to have a limit for the number of cookies, it truncates the matching cookie list if the length of the cookie
header exceeds 8 KB. We did not observe similar behaviors in other browsers.
Cookie Name. The cookie name can contain all
US-ASCII values except control characters and separator characters (see definition in [2] and [8]). We
found that Safari mistakenly stores cookie name in caseinsensitive manner. Some programming languages also
implement cookie names incorrectly. Previously Lundeen et al. reported that ASP.NET implements cookie
names case-insensitively [24]. We found that ASP makes
same mistake. In addition, PHP performs percentdecoding on cookie names. For these languages, different cookie names sent by browser are possibly recognized as same name cookies, which embraces another
vector for cookie shadowing. For example, PHP interprets a cookie header %76alue=bad; value=good; as
value=bad; value=good;, causing the good cookie to
be shadowed by the bad one.
Cookie Path. According to the specification, a cookie
matches a URL only when the path scope of the cookie is
a sub-directory of (or identical to) the URL path. When
a cookie does not specify the path scope, the browser
is required to set its default path as the directory-portion
of the URL path without any trailing slash. We found 4
violations to the standard: 1) IE and Safari4 implement
a sub-string other than sub-directory matching rule; 2)
Firefox matches cookie path with not only the URL path,
but also the URL query and the URL fragment portion;
3) Firefox matches a cookie path with a URL path when
the former has one more slash than the later; 4) Chrome,
Safari, and Opera (Linux and iOS versions) include the
trailing slash in default cookie path.

Pitfalls in Cookie Implementations

Based on the threat model and the understanding of potential attack surfaces, we then turn to understand how
cookie related mechanisms are implemented in browsers
and web applications. Our study pinpointed a number
of inconsistent and/or non-conforming behaviors in major browsers and web frameworks, as summarized in Table 4. We also identified several vulnerabilities in major browsers allowing an active network attacker to inject
cookies even when the full HSTS is deployed. We have
reported these vulnerabilities to browser vendors.

4.1

Uncovered Implementation Quirks

Browser-side Cookie Ordering. The current cookie

specification [2] suggests that browsers should rank
cookies first by path specificity and then by the creation
time in ascending order. We found all major browsers
follow this suggestion except Safari, which ranks cookies first by the specificity of the domain attribute then by
the path specificity.
Server/script-side Cookie Preference. The cookie
header is semantically a list. For the same name cookies
in the list, the specification states that the server should
not rely upon cookies ordering presented by the browser.
We examined popular web programming languages, web
frameworks, and third-party libraries including PHP,
Python, Java, Go, ASP, ASP.NET, JavaScript, Node.js,
JQuery, JSF, SpringMVC. At the language level, only
Java, JavaScript and Go provide built-in or standard library interfaces to read cookies as a list. Other languages, and all web frameworks and third-party libraries
treat the cookie list as a name-value map that only returns
one value for each cookie name in the list. For cookies

4 Also Chrome on iOS, but as iOS browsers need to use Apples

rendering engine rather than their own, this is probably due to Apples
decision, not Googles

6
712 24th USENIX Security Symposium

USENIX Association

Cookie Property
Browser-side priority
Server/script-side preference

Specification
Cookies SHOULD be ranked by specificity of
path then by creation time in ascending order.
Server SHOULD NOT rely on cookies ordering presented by browsers.

Cookie storage limitation

Several vague suggestions

Cookie header size limitation

Cookie name

Not specified
US-ASCII values except control characters
and separator characters (see definition in [2]
and [8])
1. Cookie path and URL path MUST be identical or sub-directory matching;
2. Trailing slash MUST NOT be included in
default cookie path.

Cookie path

Non-conforming/inconsistent behaviors
1. Safari ranks cookies by specificity of domain then by specificity of path.
1. Most standard libs and frameworks only provide name-value map interfaces;
2. For each name in the cookie list, Python prefers the last-ordered cookie, others
prefer the first-ordered one.
1. Safari seemingly does not have limitation on the number of stored cookies;
2. Chrome and Firefox limit the size of the cookie store per base domain, IE does
so per specific domain scope.
1. Safari truncates the cookie header not to exceed 8,192 bytes.
1. Safari is case-insensitive with cookie name;
2. ASP and ASP.NET are case-insensitive;
3. PHP performs percent-decoding on cookie name.
1. Firefox matches cookie path not only with URL path, but also with URL query
and URL fragment;
2. IE and Safari implement sub-string matching other than sub-directory matching;
3. Firefox allows cookie path has one more slash than the URL path;
4. Chrome, Safari, and Opera under some platforms include trailing slash in the
default cookie path.

Table 4: Summaries of non-conforming and inconsistent behaviors found in browser and web server cookie implementations.

4.2

Uncovered Vulnerabilities

IE
Chrome
Firefox
Safari
Opera

Vulnerabilities in Handing Proxy Response. In [6],

Chen et al. found a number of flaws in major browsers.
The root problem resided in the handling of HTTPS responses. Essentially, all browsers at that time could not
differentiate an HTTPS response from a proxy and an
HTTPS response from the intended server. The flaws
were patched after disclosure. However, we found the
patches are incomplete: if a proxy replies to a HTTPS
CONNECT request with an unencrypted 407 (proxy authentication required) response, all major browsers except IE accept the cookies set in 407 response. While
some vulnerable browsers display a pop-up window,
some accept cookies silently (Table 5).
These vulnerabilities allow a malicious proxy to
launch cookie injection attacks against a full HSTS site.
Users who use proxies or have them set automatically,
these vulnerabilities can also be exploited by an active
MITM between the victim and the proxy, even if a victim
user does not intentionally use the attacker as the proxy.
Vulnerability in Handing Public Suffixes in Safari.
As described in Section 3.2, the public suffix list enforces
the boundary between public and non-public cookie domain scopes. However we found the implementation
of Safari is vulnerable under certain conditions. When
Safari issues a request http://tld/, it accepts cookies in the response with domain scope as .tld, which
are shared by all subdomains.tld. Because HSTS is
not enabled on an entire TLD (in general, there is no A
record indicating a server at the TLDs), this vulnerability
is exploitable by active network attackers who can forge
a DNS response as well as an HTTP response.
Vulnerability in Safaris HSTS Implementation.
We also found a vulnerability in Safaris HSTS implementation. When receiving a URL, Safari does percent-

Windows






Mac OS

Linux

Android

iOS

N/A

N/A

N/A

N/A












N/A

N/A

N/A

N/A

N/A

: cookie injection with pop-up window.

: cookie injection without pop-up window.
: cookie injection and script injeciton.

Table 5: Browser vulnerabilities in handling 407 response by a malicious proxy.

decoding and upper-to-lower case conversion on its domain name before issuing a request. However, the HSTS
check is performed before the conversion process completes, enabling an attacker to bypass Safaris HSTS
check if both capital and percent-encoding are used in
the domain name.

Real-World Exploitations

Our study aims at understanding the prevalence and

severity of potential exploitation by cookie injection in
real-world websites. In particular, we are curious about
how web developers use cookies, whether they are aware
of this problem explicitly and have developed best practices accordingly. With these questions in mind, we
conducted black box penetration tests on a number of
popular websites with our test accounts. We also reviewed several well-known open source web applications. For penetration tests, we first used browser extensions like EditThisCookie [1] to test manually. For possible exploitations, we then implemented with Bro [27]
(for packet sniffing and injection with the rst tool) in an
open wireless network setting.
We found cookie injection attacks are possible with
7

USENIX Association

24th USENIX Security Symposium 713

5.1.1

very large websites and popular open source applications

including Google, Amazon, eBay, Apple, Bank of America, BitBucket, China Construction Bank, China UnionPay, JD.com, phpMyAdmin, and MediaWiki, among
others. The consequences of attacks include, but are not
limited to, XSS, privacy leakage, bypassing of cross-site
request forgery (CSRF) defenses, financial loss, and account hijacking. The varieties of vulnerable web applications and exploitations suggest cookie injection is a
serious threat in the real world, and deserves a greater
attention from the web security community.
The exploitations we found indicate three common
cookie usages: 1) using cookies as authentication tokens;
2) associating important and session independent states
with cookies; 3) reflecting cookies into HTML. These
cookie usages often lead to cookie injection attacks if
specific defensive measures are not in place.
We present our exploitations based on these categories, along with the necessary background and additional observations. Please refer Section 4 and Table 4
for the details of different cookie implementations involved in some cases. We extensively make use of cookie
shadowing. For these cases, unless otherwise specified,
we assume that the web server has the common behavior of preferring the first-ordered cookie for each name
in the cookie list.

5.1

Exploiting Google Chat and Search

We first present two exploits targeting Google, which

lead our observation of the sub-session hijacking attack.
Googles base domain google.com is not protected with
full HSTS, so in most cases it is subject to cookie injection by an active network attacker.
Case-1: Gmail chat gadget hijacking. The web
interface of Gmail at https://mail.google.com/
shows a chat gadget at the bottom left corner. If an attacker hijacks the gadget without affecting Gmail and
other Google services, he can fake the victims friend list
and chat with the victim to initiate advanced phishing,
intercept communication, or perform other disruptive activity. This could be particularly deceptive in a targeted
attack scenario.
We have confirmed this attack. Although the browser
displays everything as one page, the chat gadget and
Gmail content are actually loaded with different URLs
then composed together. Both the chat gadget and Gmail
use cookies for authentication. If an attacker injects his
Google session cookies in a way that the injected cookies
shadow the original ones only at the chat gadget related
URLs, then the attacker can put his chat gadget on the
victims screen, without disturbing the victims use of
Gmail and other Google services.
We demonstrated this attack by injecting a total of 25
cookies: five session cookies SID/SSID/HSID/APISID/
SAPISID, each with five specific paths. Meanwhile
most Google services are not affected because the specific paths of the injected cookies do not match with their
URL paths. This is sufficient to cause the chat window
to load with the attackers cookies, while all other components are loaded as the victim.
Case-2: Invisible Google search history stealing. Another attack is to use cookie shadowing to steal
Google search history (which is automatically logged
and retrievable with the login cookie) without being noticed. We assume that a user has visited https://
www.google.com/, which shows the search box and her
profile name and icon. When she types in the search
box, browser-side script issues AJAX requests to https:
//www.google.com/search to get search results.
Our original goal was to only shadow the session
cookies of the AJAX request, so that we could steal
search history without affecting the web interface loaded
by https://www.google.com/. But it turned out we
could not achieve this. We first injected three relevant
session cookies SID/SSID/HSID with path /search.
However, this attempt failed because we found the server
unusually preferred the last-ordered cookie, and the injected cookies were ranked before the legitimate ones
because of the specific path. We then found out a way
to only shadow the session cookies of the AJAX re-

Cookies as Authentication Tokens

A common practice in web development is to use a

cookie to identify a user session. Many websites further set long expiration durations on session cookies to
avoid having users sign in every time. This practice itself
is somewhat questionable, because session cookies are
sent along with HTTP requests automatically, which facilitate CSRF attacks. Nevertheless, Barth et al. showed
that CSRF attacks can be defeated with specific defensive principles and techniques in web applications [4].
Also in [4], Barth et al. noted a special form of CSRF
which they called login CSRF. In this attack, an attacker
signs in with his own account on the victims browser.
If not noticed, the victim might visit targeted web site
on behalf of the attackers account, resulting in security
and privacy consequences such as search history leakage,
credit card stealing, and XSS. The authors also pointed
out that login CSRF is a special form of a threat they
called Authenticated-as-Attacker, which can also be carried out by injecting malicious session cookies to overwrite original ones.
In fact, the consequences of cookie injection on session cookies can go beyond those described in [4]. We
found that, by using cookie shadowing, similar attacks
could be carried out without noticeable evidences by the
victim. We call our attacks sub-session hijacking attacks.
8
714 24th USENIX Security Symposium

USENIX Association

quest on Safari by exploiting its cookie header limitation (see Case-5 in Section 5.1.5 for the details). However, the server seemed to check whether session cookies under https://www.google.com/search are consistent with those under https://www.google.com/.
Once receiving inconsistent session cookies from the
AJAX request, it navigated the web interface to https:
//www.google.com/search, which still showed the attackers profile name and icon.
Our final attack was to inject session cookies with domain scope www.google.com and path /, so that for
non-Safari browsers, the attacker could steal the victims
search history. Although this attack affects the web interface, causing to show the attackers profile name and
icon, it does not affect most other Google services. We
also verified an invisible attack by spoofing the victims
profile name and icon.
5.1.2

which may be noticed by the victim. However, if the attacker can only hijack an AJAX request which is not related to the interface, especially ID-indicators, the attack
could be launched invisibly.
Second, explicit and session dependent verifiers could
bind separate URLs together, so that the attacker needs
to hijack more URLs. One example is using a session
dependent nonce to counter CSRF attacks. Suppose the
attacker wants to steal some sensitive information submitted by a form which is fetched from URL GetForm
then submitted through URL SubmitForm. If the CSRF
protection of the form is session dependent, e.g. a nonce
associated with user session embedded in the form and
verified when submitting, the attacker must hijack both
GetForm and SubmitForm so that the CSRF verification does not fail. Otherwise he only needs to hijack
SubmitForm.
It turns out that sub-session hijacking can be a powerful attack against todays websites. Because many
web applications do not adopt mechanisms to bind subsessions together, and, for many operations, hijacking
one sub-session is sufficient to cause serious consequences. Below we describe three common functionalities that are often vulnerable to sub-session hijacking,
demonstrated with real-world cases.

Sub-session Hijacking Attacks

The two cases above show a common pattern: the attacker intends to limit the effective scope of injected session cookies as small as possible to reduce the visibility
of his attack.
Essentially, web applications require one or more
request-reply pairs with different URLs, which we view
as different sub-sessions. In a normal case, when a user
views a web page or performs a certain action through
a series of pages, the corresponding sub-sessions carrying the same user authentication tokens are attributed to
the users account. However, when using cookies as authentication tokens, the cookie-URL matching rules and
implementations often allow the attacker to selectively
associate one or more sub-sessions to the attackers account by cookie shadowing. That is why we call this type
of attack sub-session hijacking attacks.
The impact of such attacks varies by the applications.
In general, the attackers strategy is to select a minimum
set of sub-sessions that achieve his attack goals meanwhile keep the visibility of the attack as small as possible. However such attack could be made difficult by
some implementation choices.
First, in general, a victim could notice a sub-session
hijacking attack if she views abnormal changes of some
visual elements on her screen. Typically such elements
include username, email, a profile icon etc., which we
refer to as ID-indicators. If a website uses less URLs in
one page or one certain functionality, and makes the important URLs related with the ID-indicators, the attacker
is less likely to perform sub-session hijacking without
being noticed. For example, in Case-2, the attacker has
to hijack both of the URL that shows the search interface,
and the AJAX request that performs the search. This limitation causes the expose of his profile name and icon,

5.1.3

Payment Account Stealing

Many websites require users to associate one or more

payment accounts to pay their bills or online purchases.
If the attacker hijacks the payment account submission
form, he could get sensitive information, or even spend
money using the victims payment account.
Case-3: Credit card stealing on China UnionPay.
China UnionPay, a government-owned financial corporation in China, has an online third-party payment service in which users can add their credit/debit cards. Although the process of adding a card involves four URLs
as well as authentication via text message, all the URLs
merely use one session cookie uc s key for authentication and the actual data submission is performed at one
URL that is not related to any ID-indicator. We have
verified that by shadowing the session cookie at the submission URL, the attacker can hijack China UnionPays
credit card association invisibly to obtain the victims
(obfuscated) credit card number and its spending history
when the victim uses this interface in the future.
5.1.4

Online Deposit Hijacking

A common feature in many Chinese websites is the ability to deposit money from an online bank (or a third-party
payment service like Alipay) into a website for future
9

USENIX Association

24th USENIX Security Symposium 715

and BitBucket. BitBucket, a popular code hosting

service, provides account association with Google by
OAuth. If a user is already logged in with Google
and has authorized BitBucket to access her Google profile through OAuth, the association is accomplished
with two forth-and-back redirections with https://
accounts.google.com/o/oauth2/auth without confirmation except a final message saying Youve successfully connected your Google account.
Our goal is to hijack the Google OAuth URL to invisibly cause an association violation. There are 5 relevant session cookies: SID/SSID/HSID with domain
scope .google.com and path /, and LSID/LSOSID
with domain scope .accounts.google.com and path /.
This is challenging because the server seemingly has
deployed specific defense to counter cookie injection.
First, accounts.google.com has enabled HSTS with
includeSubDomains. Second, if we inject cookies with
the same names, the server redirects us to a CookieMismatch warning page.
We successfully launch the attack on Safari by taking advantage of Safaris quirks. First, we exploit the
HSTS implementation bug (Section 4.2) to inject the attackers five session cookies with domain scope .accounts.google.com and path /o/oauth2/. Recall that
Safari ranks cookies by domain specificness then by path
specificness, therefore the injected cookies shadows the
legitimate ones. Then, we make use of Safaris 8 KB
limitation on the cookie header (see Section 4.1) to get
around the same name cookie detection. To achieve
this, we inject a number of cookies with specific names
and domain/path scopes, so that they are ranked between the injected session cookies and the legitimated
session cookies. We control the length of these cookies
to overflow the cookie list so that Safari truncates the
legitimated session cookies when issuing requests to the
OAuth URL. This allows us to bypassed all restrictions.

spending. We found this feature is particularly vulnerable to sub-session hijacking.

The process of online deposit usually includes six
steps: 1) the user enters deposit amount; 2) the website
generates an ID as a unique identifier of this transaction;
3) the website redirects the user to the selected online
bank with the transaction ID; 4) the user authenticates
and confirms to withdraw money from the online bank;
5) the online bank notifies the website with the transaction ID and redirects the user back to the website; 6) the
website receives the notification from the online bank,
and adds the corresponding amount on the users account
according to the transaction ID. The bank site only shows
the transaction ID on its interface which is usually an unmeaningful string. If the attacker can hijack the step 2 to
associate the transaction ID with his account without being noticed, the victim user is likely to finish all steps on
the online bank because there is no abnormal visual indication. Once the victim does so, the money is deposited
to the attackers account.
Case-4: Deposit hijacking on JD.com. JD.com is a
popular E-commerce website in China. In its implementation of the online deposit feature, the second step uses
an AJAX request that is not related to any ID-indicator.
We have verified that by shadowing JD.coms session
cookie ceshi3.com at the AJAX request, the attacker can
hijack the online deposit invisibly, redirecting funds from
the victim into the attackers jd.com account.
5.1.5

Account Hijacking in SSO Association

Single Sign On (SSO) is a technique where an Identity

Provider (IdP) provides authentication assertions for a
logged-in user to relying parties (RP) for them to authenticate the user. SSO usually enables automatic login on
the relying party, providing a better user experience and
in some cases better security. This is a popular technique
deployed by a number of large websites such as Google
and Facebook as IdPs, and many other web sites as relying parties. Popular web protocols used for SSO implementation include OpenID [10] and OAuth [14].
Under certain conditions, SSO systems face a threat
called association violation [31], in which a victim account on an RP is associated with an attackers account
on an IdP, so that the attacker gains control of the victims account on the RP. This is likely to happen when 1)
the victim is logged-in in the IdP as the attacker, 2) the
RP has a feature for its users to associate with their accounts on the IdP, 3) the feature is implemented through
redirections without further confirmation. The first condition can be mounted by cookie injection, and the websites satisfying the latter two conditions are not hard to
find.
Case-5: Account hijacking against Google OAuth

5.2

Cookies as References to Session Independent States

Session fixation is a well-known attack in which an attacker holds a session ID, then persuades a victim to authenticate with that ID so that he gains control of the
victims account [21]. Cookie injection can be used to
exploit vulnerable websites that use cookies to store session IDs. Standard defenses, e.g. regenerating session ID
after login, have been widely adopted.
However, we found that, although some websites have
implemented defenses for typical session fixation attacks, they still have similar vulnerabilities for cookie
injection. The root cause is that they associate important server-side states with long-term cookies. Moreover, they do not bind these states with user sessions.
10

716 24th USENIX Security Symposium

USENIX Association

The attacker can fixate such cookies by cookie injection (e.g., through cookie overwriting) in order to access
and manipulate the associated states. Interestingly, most
of the vulnerable websites we found vulnerable are Ecommerce websites.
Case-6: Shopping cart tracking/manipulation on
popular E-commerce websites. We demonstrate this
type of issues on 3 popular E-commerce websites:
Apple, eBay, and JD.com. These websites allow unregistered visitors to add items in shopping carts. For better
user experience, they never expire the shopping carts
on the server side, and use long-term cookies on the
browser side as references. We have verified that if the
attacker fixates the corresponding cookies using cookie
injection, he can track or manipulate shopping carts of
the unregistered visitors.

track all purchases of the victim. To do so, he first

creates an on-going purchase, of which the internal
data structure is also shared with the victim. Later,
when the victim makes a purchase, the information
is updated to the shared data structure, consequently
retrieved by the attacker. On Amazon China, the attacker can see all information of the victims purchase including items, amount, the victims name,
delivery address, and cellphone number. On Amazon U.S., the delivery address and cellphone number are not visible by the attacker.
Potential hijacking of purchases. When detecting
an ongoing purchase by the victim, the attacker can
change the delivery address so that the purchase is
paid by the victim but sent to the attacker. If the
victim confirms the hijacked purchase, she cannot
even see where the purchase is hijacked to in her
order history, because Amazon only shows Gifting address. The attacker can even manipulate the
purchase in such a way that it is paid by the victim, delivered to the attacker, and only recorded in
the attackers order history. The only limitation of
the attack is that if the delivery address is new to
the payment option, Amazon requires the victim to
confirm the card number, however the interface is
arguably not alarming. On Amazon China, this limitation does not apply if the victim chooses to pay
with a third-party service like Alipay.

We also found similar problems on Amazon, which

are much more serious in terms of the real-world
consequences, because they compromise security for
registered users.5
Case-7: Browsing history and purchase tracking/hijacking on Amazon. Amazons E-commerce
websites use two long-term cookies session-id and
ubid-main to associate with a users browsing history
and the ongoing purchase. Surprisingly, these important
states are not associated with the user session (Not as its
name suggests, session-id is not used for user authentication). Once the attacker fixates the two cookies, he
can launch various attacks remotely.
The first exploitation is to track and manipulate the
users browsing history. Amazon keeps all previous
viewed items in a users browsing history. Upon fixating the two referencing cookies, the attacker can track
what the victim have viewed on Amazon in real-time. He
can also inject unwanted items into the browsing history,
which affects the recommendation system.
Moreover, from what we observed, we infer that Amazon keeps a session independent data structure for an
on-going purchase, which stores the user, the purchased
items, the total amount, the delivery address, and other
payment information. The structure is likely created by
clicking the proceed to checkout button, and released
after clicking of the place your order button. This
structure is associated with the same two cookies referencing the browsing history. By fixating the two cookies
and consequently gaining access of the data structure, the
attacker has various ways to manipulate the victims purchase remotely. Below we describe two exploitations:

5.3

Cookies reflected into HTML

Another common practice is to store auxiliary variables

like preferred language or username as cookies, and reflect these cookies into HTML or script snippets. If not
implemented carefully, this practice could make websites
vulnerable to various attacks in face of cookie injection.
5.3.1

XSS via Cookie Injection

A direct threat is XSS: if reflected cookies are not

sanitized sufficiently, the attacker can embed malicious
scripts in reflected cookies to launch XSS attacks through
cookie injection.
Case-8: XSS via cookie injection on China Construction Bank, Amazon Cloud Drive, eBay and others. We found a number of websites do not validate reflected cookies sufficiently. Using cookie injection, we
successfully mounted XSS against China Construction
Bank, Amazon Cloud Drive, eBay and several other websites.
Case-9: Insufficient cookie validation on Bank of
America. Among the XSS vulnerabilities we found, the
one on the Bank of America website is fairly unique. We

Tracking of all purchases. First, the attacker can

5 However, we note that many E-commerce sites, including Amazon, use mixed content, and thus are also vulnerable to attackers injecting scripts into the insecure domain that remain in the browser cache.

11
USENIX Association

24th USENIX Security Symposium 717

found that one cookie with path / on Bank of Americas website could be exploited to inject XSS. At first,
our nave exploitation by overwriting the cookie with a
XSS payload failed. The limitation was that the website
performed a strict validation on the cookie at the login
URL. Only if the cookie was absent would the website
set a clean value on the cookie from the response of the
login URL, then used it in subsequent requests without
validation. Our nave exploitation was prevented by the
strict validation at the login URL.
We found a technique to bypass the limitation, so that
the XSS payload can be buried into the victims browser
and triggered when next time the victim logs in by injecting multiple cookies. We injected two cookies. The
first one had the same 3-tuple identifier as the legitimate
one, but with an expired time to ensure the legitimate
cookie was discarded and absent at the login URL. The
second injected cookie contained the XSS payload and
had a different cookie path /myaccount that matched
with the first URL after login. Although the server set a
clean cookie in the response of the login URL, the specific path of the second injected cookie not only avoid
being overwritten by the clean cookie, but also shadowed
the clean cookie in subsequent requests, triggering a successful XSS attack.
This case implies a possible misconception that performing a complete cookie validation on one entry
point is sufficient. In fact, because of the asymmetry
between cookie read and write operations, every different
URL might bring different and unexpected cookie values
no matter how server set cookies in previous responses.
Developers must treat every request as a new entry point
and carefully validate all associated cookies.
5.3.2

the second condition. If a webpage contains a reflected

cookie, the attacker can abuse it with cookie injection
as the first vector to launch the BREACH attack to infer
secrets in this webpage.
Case-10: BREACH attacks on phpMyAdmin and
MediaWiki. We found phpMyAdmin, a popular open
source web application for remote database management, reflects a cookie for language preference after
sanitization in error page if its value is invalid. The
BREACH attack using this cookie can reliably infer the
CSRF token in the error page, enabling further CSRF attacks. Similarly, MediaWiki reflects a cookie into its login form, also allowing the BREACH attack to infer the
CSRF token in the login page.

5.4

Summary

These exploitations show that cookie injection enables

undesired and complicated interactions among cookie
implementations, web applications, and various known
threats. It is clear that our empirical assessment only
touches a part of the whole problem space. Nevertheless, we believe these cases demonstrate that the security
implications of cookies lack of integrity are not well and
widely understood by the community, and current cookie
practices have widespread problems when cookie injection is taken into consideration.
Report and Response. We have reported all vulnerabilities to the affected websites. Some have acknowledged (e.g., Amazon), and some (e.g., Bank of America)
have fixed the issues.

Possible Defenses

Some existing techniques can help mitigate this threat,

including full HSTS, public suffix list, defensive cookie
practices, and anomaly detection.
Full HSTS and Public Suffix List. We strongly
recommend that websites deploy full HSTS to prevent
cookie injection from active network attackers, as this
provides complete protection once a site is pinned by a
user visit. The community should also make the effort to
raise the awareness of cookie injection attacks, and clarify the different levels of security provided by HTTPS,
HSTS, and full HSTS. For websites that host shared domains, the best practice is to use separate domains and
register them on the public suffix list. Efforts also should
be made to increase the awareness of cookie injection
from shared domains and the public suffix list.
Defensive Cookie Practices. For websites that cannot enable full HSTS, and have concerns about cookie
injection from related domains, defensive cookie practices may mitigate certain cookie injection threats. For
example, frequently invalidating session cookies could

BREACH Attacks through Cookie Injection

In 2002, Kelsey observed that when combining encryption with some compression algorithms, the size of compressed data can be used as a side channel, potentially
causing plaintext leakage under certain conditions [20].
Rizzo and Duong found a real-world case in 2012, named
as CRIME attack, in which an active network attacker
initiates encrypted HTTP requests from a victim browser
with different URLs as partially-chosen plaintexts, then
infer embedded secrets like session cookies by observing
the sizes of the compressed and encrypted requests [29].
Rizzo and Duong also mentioned that a similar attack
could also be mounted to infer secrets in encrypted
HTTP responses. This was explored and demonstrated
by Gluck et al., named as the BREACH attack [12].
BREACH requires the attacker to be able to 1) inject a
partially-chosen plaintext into the HTTP response of one
webpage, and 2) measure the size of the compressed then
encrypted response. An active network attacker satisfies
12
718 24th USENIX Security Symposium

USENIX Association

reduce the risk of sub-session hijacking. Instead of using

cookies, Websites can also use new features in HTML5
like localStorage and sessionStorage to facilitate
browser-side state management, which does not have
cookies integrity deficiencies, although these mechanisms are less convenient for cross-protocol and crossdomain state sharing.
Anomaly Detection. Websites should consider detecting same name cookies in the cookie header, as we
discussed in the accounts.google.com case. This is
reasonable because same name cookies should not be
considered a legitimate use according to both the specification and the inconsistent implementations. This detection would protect non-Safari users from attacks using
cookie shadowing.

6.1

2. Cookies with the secure flag MUST be given

higher priority over non-secure cookies.
3. A browser MUST only send the highest priority
cookie for any cookie name.
4. In removing cookies due to a too-full cookie store,
the browser MUST NOT remove a secure cookie
when there are non-secure cookies that can be removed.
5. The browser MUST allow an HTTP connection
to clear a secure cookie by setting an alreadyexpired expiration date, but the browser MUST
NOT remove the cookie from the store. Instead,
the browser MUST set the do not send flag and
maintain the original expiration date.
6. The browser MUST NOT send a cookie with the
do not send flag, nor send any non-secure cookie
with the same name.

Proposed Browser Enhancements

We propose several browser-side enhancements to mitigate cookie injection attacks. Our proposals do not require any server-side change, so they would benefit many
legacy websites.

Currently, Chrome, Firefox and Safari, but not Internet

Explorer, have deployed the HSTS support. We believe
that if all major browsers could deploy it, websites with
full HSTS would be capable of defending against active network attackers in most cases. However, deploying full HSTS needs all subdomains to support HTTPS
with valid certificates. There are a number of practical
hurdles for websites to satisfy such a strict requirement.
For example, Google cannot enable full HSTS, because
it is required to support non-HTTPS access for mandatory adult-content filtering at school and some other locations [13]. Kranch and Bonneau also reported the current incapability of Facebook and Twitter to deploy full
HSTS [22]. Hence, we believe full HSTS is not likely to
be adopted widely in the near future.
To protect a site which cannot deploy full HSTS, a
browser must not allow an HTTP connection to replace
or shadow secure cookies, effectively adding an HSTSlike pin for any secure cookie. We propose to modify the
semantics of the existing cookie store by adding a do
not send flag and changing the cookie store behavior
with the following semantics. We believe these semantic
change should provide protections while minimizing the
disruption to existing sites:

The first rule prevents an active network attacker from

injecting or replacing secure cookies. The second and
third rules combined prevent an active network attacker
from shadowing a secure cookie. The fourth rule prevents an attacker from flooding the cookie store to evict
secure cookies. The fifth and sixth rules are subtle but
necessary: mixed-content sites might have a logout
button in HTTP which clears secure session cookies.
We wish to enable this functionality without allowing attackers to remove and replace a secure cookie.
Taken together, our proposals should add HSTS-like
pinning to secure cookies within the existing cookie
store. If a cookie was set with secure flag, an active
network attacker can only delete it, which largely mitigates cookie injection attacks 6 .
Compatibility. We implemented the first three rules
as a Chrome extension7 , and used the extension to
manually examined the Alexa top 40 websites. We
found one broken case: the signing out operation on
http://www.bing.com/ results in a request-reply with
http://login.live.com/logout.srf which expires
several secure session cookies under its SSO IdP domain live.com. Allowing HTTP to clear secure cookies should improve compatibility with such signing-out
practice.
We also crawled the Alexa top 100,000 domains with
both HTTP and HTTPS. In total, 48,039 domains responded with cookies. 152 (0.32%) domains returned
secure cookies over HTTP; 570 (1.19%) domains responded with cookies that have same name yet different

1. A browser MUST NOT accept a cookie presented

in an HTTP response with the secure flag set, nor
overwrite an unexpired secure cookie, except the
case in 5.

6 The non-conforming cookie name behaviors of PHP, ASP, and

ASP.NET described in Section 4.1 still expose some possibilities for
cookie shadowing. We suggest vendors to fix these incorrect implementations.
7 https://github.com/seccookie/ExtSecureCookie

6.1.1

Mitigating Active Network Attackers

13
USENIX Association

24th USENIX Security Symposium 719

domains and/or paths. These numbers suggest secure

cookies over HTTP (incompatible with the first rule) and
same name cookies (related to the second and third rules)
are rare in real-world websites. We manually examined
10 domains in each case with our extension and did not
observe evidence of broken behaviors.
While the results from our compatibility testing are
promising, we acknowledge they are preliminary. First,
we may have missed subtle incompatibility issues in our
manual testing. Second, some incompatibility behaviors
may only occur with logged-in sessions and/or specific
paths, which our testing may have failed to uncover. We
hope our limited experiments will motivate browser vendors to conduct large-scale in-depth compatibility evaluation.
6.1.2

of integrity, its root cause, and its security implications.

However, prior understanding of the subtlety, prevalence,
and severity of this problem in the real world is limited.
We take a much closer look at the problem space, provide a number of new empirical assessments which we
believe will help the community understand the problem
more deeply and know the status quo better. Specifically,
we conduct a detailed measurement of full HSTS adoption and reveal the threat to CDN customers. Prior to our
work, Kranch and Bonneau recently studied full HSTS
deployment practice but within a different context [22].
The cookie related problems revealed in our assessment
of browser and server libraries are largely unknown, except a few fragmented knowledge from Lundeens [23]
and Lundeen et al. s work [24]. The attack cases we
present also supplement previous discussion on potential exploitations in both breadth and depth. Our closeup study also leads us to find promising cookie isolation
enhancements that only require browser-side adoption.
In contrast, the previous proposed defenses need both
browser- and server-side changes [4, 5].
Broadly, our work can be viewed as an in-depth case
study of inconsistent access control policies in web.
Jackson and Barths [17] and Singh et al. s work [30]
explored this general problem, and each provided various instances. One example illustrated by Jackson and
Barth is the ability of JavaScript to read all cookies with
matching domain scopes regardless of their paths [17].
This behavior has now been noted explicitly in the current specification [2].
Security Related Cookie Measurement. Zhou and
Evans studied the rare deployment of the HTTPOnly
cookies at the time [33]. They believed that the requirement of both client and server changes played an important hurdle in its adoption. Kranch and Bonneau found
many websites deploy HSTS yet do not marked their
cookies with the secure flag, which are vulnerable to
cookie theft in certain conditions [22]. These two measurements were concerned with cookies confidentiality,
while our work looks at the other property, i.e. cookies
integrity. Singh et al. measured the real-world usages
of secure cookies (0.07%, 62 out of 89,222 sites) over
HTTP and same name cookies (they called duplicate
cookies) (5.48%, 4,893 out of 89,222 sites) [30]. Our
assessment obtains similar results.

Mitigating Web attackers

A domain can set cookies with a more specific domain

scope (e.g. host-only) to prevent cookie stealing by XSS
from sibling domains. But this currently has no effect
on cookie injection since injected cookies with shared
domain scopes yet longer paths are effective for cookie
shadowing, and longer paths are available in most cases
if an adversary is in control of a related domain. Combining the second rule of the above proposals, we suggest:
7. When issuing a request, the browser MUST rank the
cookie list by a) presence of the secure flag, and b)
specificity of the domain scope.
Together with the third rule presented above, this
should enable developers to prevent cookies from being
overwritten or shadowed by using specific domain scope
(together with the secure flag when using HTTPS). We
have also implemented this policy in the same Chrome
extension mentioned above.

Related Work

Comparison to Previous Work. We are aware of several research papers that are directly related to cookies
weak SOP and integrity problem [4, 30, 24, 23, 5, 2,
32], and some other papers that are comparable to our
work [17, 22]. Among the directly related research,
Barths [2] and Zalewskis work [32] focused on explaining the cause of the cookie integrity problem. Most previous research only briefly touched cookie integrity as
a relevant subproblem rather than main topic [4, 30, 24,
23]. Bortz et al. s research is close to ours. Especially,
they introduced the notion of a related domain attacker
which we use throughout this paper. However, their work
is limited to high-level discussion [5]. In summary, previous research has discussed the problem of cookies lack

Conclusions

Cookies lack integrity. Although long known in community lore, the community has under-appreciated the implications. We have attempted to systematically evaluate
the implications of cookie integrity, including evaluating
weaknesses and evaluation artifacts in both browser and
server libraries, building real-world attacks against ma14

720 24th USENIX Security Symposium

USENIX Association

jor sites including Google and Bank of America, including subtle account-hijack attacks and XSS attacks buried
in injected cookies, and developing an alternate browser
cookie policy that mitigates the threat from networklevel attackers. We expect our work to raise the awareness of the problem, and to provide a context for further
discussion among researchers, developers and vendors.

[12] G LUCK , Y., H ARRIS , N., AND P RADO , A. BREACH: Reviving the CRIME Attack.
http://breachattack.com/
resources/BREACH%20-%20SSL,%20gone%20in%2030%
20seconds.pdf, 2013. [accessed Feb-2015].
[13] G OOGLE S UPPORT. Block Adult Content at Your School.
https://support.google.com/websearch/answer/
186669?hl=en. [accessed Feb-2015].
[14] H ARDT, D. The OAuth 2.0 Authorization Framework. IETF
RFC 6749 (2012).
[15] H ODGES , J., JACKSON , C., AND BARTH , A. Http Strict Transport Security (HSTS). IETF RFC 6797 (2012).

Acknowledgements

[16] IEB LOG. Project Spartan and the Windows 10 January Preview
Build. http://blogs.msdn.com/b/ie/archive/2015/01/
22/project-spartan-and-the-windows-10-januarypreview-build.aspx. [accessed Feb-2015].

We would like to thank our shepherd Hovav Shacham,

and the anonymous reviewers for their insightful comments. We are grateful to Vern Paxson, Frank Li and
David Fifield for valuable discussion, and Jianjun Chen
for help of some exploitations. We also thank Chris
Evans, Joel Weinberger, Chris Palmer, and Nick Sullivan for valuable feedback. This work is partially supported by the National Natural Science Foundation of
China (Grant No. 61472215), Tsinghua National Laboratory for Information Science and Technology (TNList)
Academic Exchange Foundation, and the Natinoal Science Foundation (CNS-1213157 and CNS-1237265).

[20] K ELSEY, J. Compression and Information Leakage of Plaintext.

In Fast Software Encryption (2002), Springer, pp. 263276.

References

[21] KOL S EK , M. Session Fixation Vulnerability in Web-based

Applications.
http://www.acros.si/papers/session_
fixation.pdf, 2002. [accessed Feb-2015].

[17] JACKSON , C., AND BARTH , A. Beware of Finer-Grained Origins. Proceedings of 2th W2SP (2008).
[18] JACKSON , C., AND BARTH , A. ForceHTTPS: Protecting HighSecurity Web Sites from Network Attacks. In Proceedings of the
17th WWW (2008), ACM, pp. 525534.
[19] J OHNSTON , P., AND M OORE , R. Multiple Browser Cookie
Injection Vulnerabilities. http://www.westpoint.ltd.uk/
advisories/wp-04-0001.txt, 2004. [accessed Feb-2015].

[1] Edit This Cookie. http://www.editthiscookie.com/. [accessed Feb-2015].

[22] K RANCH , M., AND B ONNEAU , J. Upgrading HTTPS in mid-air:

An empirical study of strict transport security and key pinning. In
Proceedings of the 22th NDSS (2015).

[2] BARTH , A. HTTP State Management Mechanism. IETF RFC

6265 (2011).

[23] L UNDEEN , R. The Deputies Are still Confused. Blackhat EU

(2013).

[3] BARTH , A. The Web Origin Concept. IETF RFC 6454 (2011).

[24] L UNDEEN , R., O U , J., AND R HODES , T. New Ways Im Going

to Hack Your Web App. Blackhat AD (2011).

[4] BARTH , A., JACKSON , C., AND M ITCHELL , J. C. Robust Defenses for Cross-Site Request Forgery. In Proceedings of the 15th
CCS (2008), ACM, pp. 7588.

[25] M OZZILA. Public Suffix List. https://publicsuffix.org/.

[accessed Feb-2015].

[5] B ORTZ , A., BARTH , A., AND C ZESKIS , A. Origin Cookies:

Session Integrity for Web Applications. Web 2.0 Security and
Privacy (W2SP) (2011).

[26] N GINX. Module ngx http core module. http://nginx.org/

en/docs/http/ngx_http_core_module.html#large_
client_header_buffers. [accessed Jun-2015].

[6] C HEN , S., M AO , Z., WANG , Y.-M., AND Z HANG , M. PrettyBad-Proxy: An Overlooked Adversary in Browsers HTTPS Deployments. In Proceedings of the 30th IEEE S&P (Oakland)
(2009), IEEE, pp. 347359.

[27] PAXSON , V. Bro: A System for Detecting Network Intruders in

Real-Time. Computer networks 31, 23 (1999), 24352463.
[28] P ROJECTS , T. C. HTTP Strict Transport Security. http://www.
chromium.org/hsts. [accessed Feb-2015].

[7] E VANS , C. Cookie Forcing. http://scarybeastsecurity.

blogspot.com/2008/11/cookie-forcing.html, 2008. [accessed Feb-2015].

[29] R IZZO , J., AND D UONG , T. The CRIME Attack. In EKOparty

Security Conference (2012), vol. 2012.
[30] S INGH , K., M OSHCHUK , A., WANG , H. J., AND L EE , W.
On the Incoherencies in Web Browser Access Control Policies.
In Proceedings of the 31th IEEE S&P (Oakland) (2010), IEEE,
pp. 463478.

[8] F IELDING , R., G ETTYS , J., M OGUL , J., F RYSTYK , H., M AS INTER , L., L EACH , P., AND B ERNERS -L EE , T. Hypertext
Transfer ProtocolHTTP/1.1. IETF RFC 2616 (1999).

[31] WANG , R., Z HOU , Y., C HEN , S., Q ADEER , S., E VANS , D.,
AND G UREVICH , Y. Explicating SDKs: Uncovering Assumptions Underlying Secure Authentication and Authorization. In
USENIX Security (2013), pp. 399314.

[9] F IELDING , R., AND R ESCHKE , J. Hypertext Transfer Protocol (HTTP/1.1): Message Syntax and Routing. IETF RFC 7230
(2014).
OpenID Authentication 2.0 - Final.
[10] F OUNDATION , O.
http://openid.net/specs/openid-authentication2_0.html. [accessed Feb-2015].

[32] Z ALEWSKI , M. The Tangled Web: A Guide to Securing Modern

Web Applications. No Starch Press, 2012.
[33] Z HOU , Y., AND E VANS , D. Why Arent HTTP-only Cookies
More Widely Deployed. Proceedings of 4th W2SP 2 (2010).

[11] G IT H UB. Yummy Cookies across Domains. https://github.

com/blog/1466-yummy-cookies-across-domains, 2013.
[accessed Feb-2015].

15
USENIX Association

24th USENIX Security Symposium 721