Christian Heimes \<firstname.lastname@example.org>
The results of an attack on a vulnerable XML library can be fairly dramatic. With just a few hundred Bytes of XML data an attacker can occupy several Gigabytes of memory within seconds. An attacker can also keep CPUs busy for a long time with a small to medium size request. Under some circumstances it is even possible to access local files on your server, to circumvent a firewall, or to abuse services to rebound attacks to third parties.
The attacks use and abuse less common features of XML and its parsers.
The majority of developers are unacquainted with features such as
processing instructions and entity expansions that XML inherited from
SGML. At best they know about
<!DOCTYPE> from experience with HTML but
they are not aware that a document type definition (DTD) can generate an
HTTP request or load a file from the file system.
None of the issues is new. They have been known for a long time. Billion laughs was first reported in 2003. Nevertheless some XML libraries and applications are still vulnerable and even heavy users of XML are surprised by these features. It's hard to say whom to blame for the situation. It's too short sighted to shift all blame on XML parsers and XML libraries for using insecure default settings. After all they properly implement XML specifications. Application developers must not rely that a library is always configured for security and potential harmful data by default.
Table of Contents
The Billion Laughs
attack -- also known as exponential entity expansion --uses multiple
levels of nested entities. The original example uses 9 levels of 10
expansions in each level to expand the string
lol to a string of 3 *
10 9 bytes, hence the name "billion laughs". The resulting
string occupies 3 GB (2.79 GiB) of memory; intermediate strings require
additional memory. Because most parsers don't cache the intermediate
step for every expansion it is repeated over and over again. It
increases the CPU load even more.
An XML document of just a few hundred bytes can disrupt all services on a machine within seconds.
A quadratic blowup attack is similar to a Billion Laughs attack; it abuses entity expansion, too. Instead of nested entities it repeats one large entity with a couple of thousand chars over and over again. The attack isn't as efficient as the exponential case but it avoids triggering countermeasures of parsers against heavily nested entities. Some parsers limit the depth and breadth of a single entity but not the total amount of expanded text throughout an entire XML document.
A medium-sized XML document with a couple of hundred kilobytes can require a couple of hundred MB to several GB of memory. When the attack is combined with some level of nested expansion an attacker is able to achieve a higher ratio of success.
Entity declarations can contain more than just text for replacement.
They can also point to external resources by public identifiers or
system identifiers. System identifiers are standard URIs. When the URI
is a URL (e.g. a
http:// locator) some parsers download the resource
from the remote location and embed them into the XML document verbatim.
Simple example of a parsed external entity:
The case of parsed external entities works only for valid XML content.
The XML standard also supports unparsed external entities with a
External entity expansion opens the door to plenty of exploits. An attacker can abuse a vulnerable XML library and application to rebound and forward network requests with the IP address of the server. It highly depends on the parser and the application what kind of exploit is possible. For example:
External entities with references to local files are a sub-case of
external entity expansion. It's listed as an extra attack because it
deserves extra attention. Some XML libraries such as lxml disable
network access by default but still allow entity expansion with local
file access by default. Local files are either referenced with a
file:// URL or by a file path (either relative or absolute).
An attacker may be able to access and download all files that can be read by the application process. This may include critical configuration files, too.
This case is similar to external entity expansion, too. Some XML libraries like Python's xml.dom.pulldom retrieve document type definitions from remote or local locations. Several attack scenarios from the external entity case apply to this issue as well.
<html> <head/> <body>text</body> </html>
|billion laughs||True||True||True||True||True||False (1)||False (5)|
|quadratic blowup||True||True||True||True||True||True||False (5)|
|external entity expansion (remote)||True||False (3)||False (4)||True||false||False (1)||False (5)|
|external entity expansion (local file)||True||False (3)||False (4)||True||false||True||False (5)|
|DTD retrieval||True||False||False||True||false||False (1)||False|
|gzip bomb||False||False||False||False||True||partly (2)||False|
|xpath support (7)||False||False||False||False||False||True||False|
|xsl(t) support (7)||False||False||False||False||False||True||False|
|xinclude support (7)||False||True (6)||False||False||False||True (6)||True|
vulnerabilities and features
disables external entity expansion
the option is ignored and doesn't modify any functionality
The defusedxml package (defusedxml on PyPI) contains several Python-only workarounds and fixes for denial of service and other vulnerabilities in Python's XML libraries. In order to benefit from the protection you just have to import and use the listed functions / classes from the right defusedxml module instead of the original module. Merely defusedxml.xmlrpc is implemented as monkey patch.
from xml.etree.ElementTree import parse et = parse(xmlfile)
alter code to:
from defusedxml.ElementTree import parse et = parse(xmlfile)
Additionally the package has an untested function to monkey patch
all stdlib modules with
All functions and parser classes accept three additional keyword arguments. They return either the same objects as the original functions or compatible subclasses.
forbid_dtd (default: False)
disallow XML with a
<!DOCTYPE> processing instruction and raise a
DTDForbidden exception when a DTD processing instruction is found.
forbid_entities (default: True)
disallow XML with
<!ENTITY> declarations inside the DTD and raise
an EntitiesForbidden exception when an entity is declared.
forbid_external (default: True)
disallow any access to remote or local resources in external entities or DTD and raising an ExternalReferenceForbidden exception when a DTD or entity references an external resource.
DefusedXmlException, DTDForbidden, EntitiesForbidden, ExternalReferenceForbidden, NotSupportedError
defusedxml.cElementTree is deprecated and will be removed in
a future release. Import from
parse(), iterparse(), fromstring(), XMLParser
parse(), iterparse(), fromstring(), XMLParser
parse(), parseString(), make_parser()
parse(), parseString(), DefusedExpatBuilder, DefusedExpatBuilderNS
The fix is implemented as monkey patch for the stdlib's xmlrpc package (3.x) or xmlrpclib module (2.x). The function monkey_patch() enables the fixes, unmonkey_patch() removes the patch and puts the code in its former state.
The monkey patch protects against XML related attacks as well as decompression bombs and excessively large requests or responses. The default setting is 30 MB for requests, responses and gzip decompression. You can modify the default by changing the module variable MAX_DATA. A value of -1 disables the limit.
DEPRECATED The module is deprecated and will be removed in a future release.
The module acts as an example how you could protect code that uses lxml.etree. It implements a custom Element class that filters out Entity instances, a custom parser factory and a thread local storage for parser instances. It also has a check_docinfo() function which inspects a tree for internal or external DTDs and entity declarations. In order to check for entities lxml > 3.0 is required.
parse(), fromstring() RestrictedElement, GlobalParserTLS, getDefaultParser(), check_docinfo()
The defusedexpat package (defusedexpat on PyPI) comes with binary extensions and a modified expat library instead of the standard expat parser. It's basically a stand-alone version of the patches for Python's standard library C extensions.
XML_BOMB_PROTECTION XML_DEFAULT_MAX_ENTITY_INDIRECTIONS XML_DEFAULT_MAX_ENTITY_EXPANSIONS XML_DEFAULT_RESET_DTD
new XML_FeatureEnum members:
XML_FEATURE_MAX_ENTITY_INDIRECTIONS XML_FEATURE_MAX_ENTITY_EXPANSIONS XML_FEATURE_IGNORE_DTD
new XML_Error members:
new API functions:
int XML_GetFeature(XML_Parser parser, enum XML_FeatureEnum feature, long *value); int XML_SetFeature(XML_Parser parser, enum XML_FeatureEnum feature, long value); int XML_GetFeatureDefault(enum XML_FeatureEnum feature, long *value); int XML_SetFeatureDefault(enum XML_FeatureEnum feature, long value);
Limit the amount of indirections that are allowed to occur during the expansion of a nested entity. A counter starts when an entity reference is encountered. It resets after the entity is fully expanded. The limit protects the parser against exponential entity expansion attacks (aka billion laughs attack). When the limit is exceeded the parser stops and fails with XML_ERROR_ENTITY_INDIRECTIONS. A value of 0 disables the protection.
0 .. UINT_MAX
Limit the total length of all entity expansions throughout the entire document. The lengths of all entities are accumulated in a parser variable. The setting protects against quadratic blowup attacks (lots of expansions of a large entity declaration). When the sum of all entities exceeds the limit, the parser stops and fails with XML_ERROR_ENTITY_EXPANSION. A value of 0 disables the protection.
0 .. UINT_MAX
Reset all DTD information after the \<!DOCTYPE> block has been parsed. When the flag is set (default: false) all DTD information after the endDoctypeDeclHandler has been called. The flag can be set inside the endDoctypeDeclHandler. Without DTD information any entity reference in the document body leads to XML_ERROR_UNDEFINED_ENTITY.
(based on Brad Hill's Attacking XML Security)
XML, XML parsers and processing libraries have more features and possible issue that could lead to DoS vulnerabilities or security exploits in applications. I have compiled an incomplete list of theoretical issues that need further research and more attention. The list is deliberately pessimistic and a bit paranoid, too. It contains things that might go wrong under daffy circumstances.
XML parsers may use an algorithm with quadratic runtime O(n 2) to handle attributes and namespaces. If it uses hash tables (dictionaries) to store attributes and namespaces the implementation may be vulnerable to hash collision attacks, thus reducing the performance to O(n 2) again. In either case an attacker is able to forge a denial of service attack with an XML document that contains thousands upon thousands of attributes in a single node.
I haven't researched yet if expat, pyexpat or libxml2 are vulnerable.
The issue of decompression bombs (aka ZIP bomb) apply to all XML libraries that can parse compressed XML stream like gzipped HTTP streams or LZMA-ed files. For an attacker it can reduce the amount of transmitted data by three magnitudes or more. Gzip is able to compress 1 GiB zeros to roughly 1 MB, lzma is even better:
$ dd if=/dev/zero bs=1M count=1024 | gzip > zeros.gz $ dd if=/dev/zero bs=1M count=1024 | lzma -z > zeros.xy $ ls -sh zeros.* 1020K zeros.gz 148K zeros.xy
None of Python's standard XML libraries decompress streams except for
xmlrpclib. The module is vulnerable
\<https://bugs.python.org/issue16043> to decompression bombs.
lxml can load and process compressed data through libxml2 transparently. libxml2 can handle even very large blobs of compressed data efficiently without using too much memory. But it doesn't protect applications from decompression bombs. A carefully written SAX or iterparse-like approach can be safe.
may impose more threats for XML processing. It depends if and how a processor handles processing instructions. The issue of URL retrieval with network or local file access apply to processing instructions, too.
DTD has more
<!NOTATION>. I haven't researched how these features may
be a security threat.
XPath statements may introduce DoS vulnerabilities. Code should never execute queries from untrusted sources. An attacker may also be able to create an XML document that makes certain XPath queries costly or resource hungry.
XPath injeciton attacks pretty much work like SQL injection attacks. Arguments to XPath queries must be quoted and validated properly, especially when they are taken from the user. The page Avoid the dangers of XPath injection list some ramifications of XPath injections.
Python's standard library doesn't have XPath support. Lxml supports parameterized XPath queries which does proper quoting. You just have to use its xpath() method correctly:
# DON'T tree.xpath("/tag[@id='%s']" % value) # instead do tree.xpath("/tag[@id=$tagid]", tagid=name)
XML Inclusion is another way to load and include external files:
<root xmlns:xi="http://www.w3.org/2001/XInclude"> <xi:include href="filename.txt" parse="text" /> </root>
This feature should be disabled when XML files from an untrusted source are processed. Some Python XML libraries and libxml2 support XInclude but don't have an option to sandbox inclusion and limit it to allowed directories.
A validating XML parser may download schema files from the information
<ead xmlns="urn:isbn:1-931666-22-9" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xsi:schemaLocation="urn:isbn:1-931666-22-9 http://www.loc.gov/ead/ead.xsd"> </ead>
You should keep in mind that XSLT is a Turing complete language. Never process XSLT code from unknown or untrusted source! XSLT processors may allow you to interact with external resources in ways you can't even imagine. Some processors even support extensions that allow read/write access to file system, access to JRE objects or scripting with Jython.
Example from Attacking XML Security for Xalan-J:
<xsl:stylesheet version="1.0" xmlns:xsl="http://www.w3.org/1999/XSL/Transform" xmlns:rt="http://xml.apache.org/xalan/java/java.lang.Runtime" xmlns:ob="http://xml.apache.org/xalan/java/java.lang.Object" exclude-result-prefixes= "rt ob"> <xsl:template match="/"> <xsl:variable name="runtimeObject" select="rt:getRuntime()"/> <xsl:variable name="command" select="rt:exec($runtimeObject, 'c:\Windows\system32\cmd.exe')"/> <xsl:variable name="commandAsString" select="ob:toString($command)"/> <xsl:value-of select="$commandAsString"/> </xsl:template> </xsl:stylesheet>
Unrestricted entity expansion induces DoS vulnerabilities in Python XML libraries (XML bomb)
External entity expansion in Python XML libraries inflicts potential security flaws and DoS vulnerabilities
Several other programming languages and frameworks are vulnerable as well. A couple of them are affected by the fact that libxml2 up to 2.9.0 has no protection against quadratic blowup attacks. Most of them have potential dangerous default settings for entity expansion and external entities, too.
Perl's XML::Simple is vulnerable to quadratic entity expansion and external entity expansion (both local and remote).
Ruby's REXML document parser is vulnerable to entity expansion attacks (both quadratic and exponential) but it doesn't do external entity expansion by default. In order to counteract entity expansion you have to disable the feature:
REXML::Document.entity_expansion_limit = 0
libxml-ruby and hpricot don't expand entities in their default configuration.
PHP's SimpleXML API is vulnerable to quadratic entity expansion and
loads entities from local and remote resources. The option
LIBXML_NONET disables network access but still allows local file
LIBXML_NOENT seems to have no effect on entity expansion in
Information in XML DoS and Defenses (MSDN) suggest that .NET is vulnerable with its default settings. The article contains code snippets how to create a secure XML reader:
XmlReaderSettings settings = new XmlReaderSettings(); settings.ProhibitDtd = false; settings.MaxCharactersFromEntities = 1024; settings.XmlResolver = null; XmlReader reader = XmlReader.Create(stream, settings);
Untested. The documentation of Xerces and its Xerces
sounds like Xerces is also vulnerable to billion laugh attacks with its
default settings. It also does entity resolving when an
org.xml.sax.EntityResolver is configured. I'm not yet sure about the
default setting here.
Java specialists suggest to have a custom builder factory:
DocumentBuilderFactory builderFactory = DocumentBuilderFactory.newInstance(); builderFactory.setXIncludeAware(False); builderFactory.setExpandEntityReferences(False); builderFactory.setFeature(XMLConstants.FEATURE_SECURE_PROCESSING, True); # either builderFactory.setFeature("http://apache.org/xml/features/disallow-doctype-decl", True); # or if you need DTDs builderFactory.setFeature("http://xml.org/sax/features/external-general-entities", False); builderFactory.setFeature("http://xml.org/sax/features/external-parameter-entities", False); builderFactory.setFeature("http://apache.org/xml/features/nonvalidating/load-external-dtd", False); builderFactory.setFeature("http://apache.org/xml/features/nonvalidating/load-dtd-grammar", False);
Copyright (c) 2013-2017 by Christian Heimes \<email@example.com>
Licensed to PSF under a Contributor Agreement.
See https://www.python.org/psf/license for licensing details.
Brett Cannon (Python Core developer)
review and code cleanup
Antoine Pitrou (Python Core developer)
Aaron Patterson, Ben Murphy and Michael Koziarski (Ruby community)
Many thanks to Aaron, Ben and Michael from the Ruby community for their report and assistance.
Thierry Carrez (OpenStack)
Many thanks to Thierry for his report to the Python Security Response Team on behalf of the OpenStack security team.
Carl Meyer (Django)
Many thanks to Carl for his report to PSRT on behalf of the Django security team.
Daniel Veillard (libxml2)
Many thanks to Daniel for his insight and assistance with libxml2.
semantics GmbH (https://www.semantics.de/)
Many thanks to my employer semantics for letting me work on the issue during working hours as part of semantics's open source initiative.
ParseErrorexception is now the same class object as
Release date: 4-Mar-2021
Release date: 12-Jan-2021
xml.etreemodule after patching
Release date: 04-May-2020
defusedxml.cElementTreeis not available with Python 3.9.
Release date: 17-Apr-2019
Release date: 14-Apr-2019
Release date: 07-Feb-2017
Release date: 28-Jan-2017
Release date: 28-Mar-2013
Release date: 25-Feb-2013
Release date: 19-Feb-2013
Release date: 15-Feb-2013
Release date: 08-Feb-2013