Public Says

Inspired by James Kettle's great OWASP AppSec Europe talk on CORS misconfigurations, we decided to fiddle around with CORS security issues a bit. We were curious how many websites out there are actually vulnerable because of dynamically generated or misconfigured CORS headers.

The issue: CORS misconfiguration

Cross-Origin Resource Sharing (CORS) is a technique to punch holes into the Same-Origin Policy (SOP) – on purpose. It enables web servers to explicitly allow cross-site access to a certain resource by returning an Access-Control-Allow-Origin (ACAO) header. Sometimes, the value is even dynamically generated based on user-input such as the Origin header send by the browser. If misconfigured, an unintended website can access the resource. Furthermore, if the Access-Control-Allow-Credentials (ACAC) server header is set, an attacker can potentially leak sensitive information from a logged in user – which is almost as bad as XSS on the actual website. Below is a list of CORS misconfigurations which can potentially be exploited. For more technical details on the issues read the this fine blogpost.

Misconfiguation	Description
Developer backdoor	Insecure developer/debug origins like JSFiddler CodePen are allowed to access the resource
Origin reflection	The origin is simply echoed in ACAO header, any site is allowed to access the resource
Null misconfiguration	Any site is allowed access by forcing the null origin via a sandboxed iframe
Pre-domain wildcard	notdomain.com is allowed access, which can simply be registered by the attacker
Post-domain wildcard	domain.com.evil.com is allowed access, can be simply be set up by the attacker
Subdomains allowed	sub.domain.com allowed access, exploitable if the attacker finds XSS in any subdomain
Non-SSL sites allowed	An HTTP origin is allowed access to a HTTPS resource, allows MitM to break encryption
Invalid CORS header	Wrong use of wildcard or multiple origins,not a security problem but should be fixed

The tool: CORStest

Testing for such vulnerabilities can easily be done with curl(1). To support some more options like, for example, parallelization we wrote CORStest, a simple Python based CORS misconfiguration checker. It takes a text file containing a list of domain names or URLs to check for misconfigurations as input and supports some further options:

usage: corstest.py [arguments] infile

positional arguments:
  infile         File with domain or URL list

optional arguments:
  -h, --help     show this help message and exit
  -c name=value  Send cookie with all requests
  -p processes   multiprocessing (default: 32)
  -s             always force ssl/tls requests
  -q             quiet, allow-credentials only
  -v             produce a more verbose output

CORStest can detect potential vulnerabilities by sending various Origin request headers and checking for the Access-Control-Allow-Origin response. An example for those of the Alexa top 750 websites which allow credentials for CORS requests is given below.

Evaluation with Alexa top 1 Million websites

To evaluate – on a larger scale – how many sites actually have wide-open CORS configurations we did run CORStest on the Alexa top 1 million sites:

$ git clone https://github.com/RUB-NDS/CORStest.git && cd cors/
$ wget -q http://s3.amazonaws.com/alexa-static/top-1m.csv.zip
$ unzip top-1m.csv.zip
$ awk -F, '{print $2}' top-1m.csv > alexa.txt
$ ./corstest.py alexa.txt

This test took about 14 hours on a decent connection and revealed the following results:

Only 29,514 websites (about 3%) actually supported CORS on their main page (aka. responded with Access-Control-Allow-Origin). Of course, many sites such as Google do only enable CORS headers for certain resources, not directly on their landing page. We could have crawled all websites (including subdomains) and fed the input to CORStest. However, this would have taken a long time and for statistics, our quick & dirty approach should still be fine. Furthermore it must be noted that the test was only performed with GET requests (without any CORS preflight) to the http:// version of websites (with redirects followed). Note that just because a website, for example, reflects the origin header it is not necessarily vulnerable. The context matters; such a configuration can be totally fine for a public sites or API endpoints intended to be accessible by everyone. It can be disastrous for payment sites or social media platforms. Furthermore, to be actually exploitable the Access-Control-Allow-Credentials: true (ACAC) header must be set. Therefore we repeated the test, this time limited to sites that return this header (see CORStest -q flag):

$ ./corstest.py -q alexa.txt

This revealed even worse results - almost half of the websites supporting ACAO and ACAC headers contained a CORS misconfigurations that could be exploited directly by a web attacker (developer backdoor, origin reflection, null misconfig, pre-/post-domain wildcard):

The Impact: SOP/SSL bypass on payment and taxpayer sites

Note that not all tested websites actually were exploitable. Some contained only public data and some others - such as Bitbucket - had CORS enabled for their main page but not for subpages containing user data. Manually testing the sites, we found to be vulnerable:

• A dozen of online banking, bitcoin and other payment sites; one of them allowed us to create a test account so we were able to write proof-of-concept code which could actually have been used to steal money
• Hundred of online shops/e-commerce sites and a bunch of hotel/flight booking sites
• Various social networks and misc sites which allow users to log in and communicate
• One US state's tax filing website (however, this one was exploitable by a MitM only)

We informed all sites we manually tested and found to be vulnerable. A simple exploit code example when logged into a website with CORS origin reflection is given below.

Leak it!

The Reason: Copy & Paste and broken frameworks

We were further interested in reasons for CORS misconfigurations. Particularly we wanted to learn if there is a correlation between applied technology and misconfiguration. Therefore we used WhatWeb to fingerprint the web technologies for all vulnerable sites. CORS is usually enabled either directly in the HTTP server configuration or by the web application/framework. While we could not identify a single major cause for CORS misconfigurations, we found various potential reasons. A majority of dangerous Access-Control-* headers had probably been introduced by developers, others however are based on bugs and bad practices in some products. Insights follow:

• Various websites return invalid CORS headers; besides wrong use of wildcards such as *.domain.com, ACAO headers which contain multiple origins can often be found; Other examples of invalid - but quite creative - ACAO values we observed are: self, true, false, undefined, None, 0, (null), domain, origin, SAMEORIGIN
• Rack::Cors, the de facto standard library to enable CORS for Ruby on Rails maps origins '' or origins '*' into reflecting arbitrary origins; this is dangerous, because developers would think that '' allows nothing and '*' behaves according to the spec: mostly harmless because it cannot be used to make to make 'credentialed' requests; this config error leads to origin reflection with ACAC headers on about a hundred of the tested and vulnerable websites
• A majority of websites which allow a http origin to CORS access a https resource are run on IIS; this seems to be no bug in IIS itself but rather caused by bad advises found on the Internet
• nginx is the winner when it comes serving websites with origin reflections; again, this is not an issue of nginx but of dangerous configs copied from "Stackoverflow; same problem for Phusion Passenger
• The null ACAO value may be based on programming languages that simply return null if no value is given (we haven't found any specific framework though); another explanation is that 'CORS in Action', a popular book on CORS, contains various examples with code such as var originWhitelist = ['null', ...], which could be misinterpreted by developers as safe
• If CORS is enabled in the crVCL PHP Framework, it adds ACAC and ACAO headers for a configured domain. Unfortunatelly, it also introduces a post-domain and pre-subdomain wildcard vulnerability: sub.domain.com.evil.com
• All sites that are based on "Solo Build It!" (scam?) respond with: Access-Control-Allow-Origin: http://sbiapps.sitesell.com
• Some sites have :// or // as fixed ACAO values. How should browsers deal with this? Inconsistent at least! Firefox, Chrome, Safari and Opera allow arbitrary origins while IE and Edge deny all origins.

To guarantee confidentiality, PDF files can be encrypted. This enables the secure transfer and storing of sensitive documents without any further protection mechanisms.

The key management between the sender and recipient may be password based (the recipient must know the password used by the sender, or it must be transferred to them through a secure channel) or public key based (i.e., the sender knows the X.509 certificate of the recipient).

In this research, we analyze the security of encrypted PDF files and show how an attacker can exfiltrate the content without having the corresponding keys.

Exfiltration via PDF Forms (B1)

Using CBC gadgets, encrypted plaintext can be prefixed with one or more chosen plaintext blocks. An attacker can construct URLs in the encrypted PDF document that contain the plaintext to exfiltrate. This attack is similar to the exfiltration hyperlink attack (A2). However, it does not require the setting of a "base" URI in plaintext to achieve exfiltration.

The same limitations described for direct exfiltration based on links (A2) apply. Additionally, the constructed URL contains random bytes from the gadgeting process, which may prevent the exfiltration in some cases.

While CBC gadgets are generally restricted to the block size of the underlying block cipher – and more specifically the length of the known plaintext, in this case, 12 bytes – longer chosen plaintexts can be constructed using compression. Deflate compression, which is available as a filter for PDF streams, allows writing both uncompressed and compressed segments into the same stream. The compressed segments can reference back to the uncompressed segments and achieve the repetition of byte strings from these segments. These backreferences allow us to construct longer continuous plaintext blocks than CBC gadgets would typically allow for. Naturally, the first uncompressed occurrence of a byte string still appears in the decompressed result. Additionally, if the compressed stream is constructed using gadgets, each gadget generates 20 random bytes that appear in the decompressed stream. A non-trivial obstacle is to keep the PDF viewer from interpreting these fragments in the decompressed stream. While hiding the fragments in comments is possible, PDF comments are single-line and are thus susceptible to newline characters in the random bytes. Therefore, in reality, the length of constructed compressed plaintexts is limited.

To deal with this caveat, an attacker can use ObjectStreams which allow the storage of arbitrary objects inside a stream. The attacker uses an object stream to define new objects using CBC gadgets. An object stream always starts with a header of space-separated integers which define the object number and the byte offset of the object inside the stream. The dictionary of an object stream contains the key First which defines the byte offset of the first object inside the stream. An attacker can use this value to create a comment of arbitrary size by setting it to the first byte after their comment.

Using compression has the additional advantage that compressed, encrypted plaintexts from the original document can be embedded into the modified object. As PDF applications often create compressed streams, these can be incorporated into the attacker-created compressed object and will therefore be decompressed by the PDF applications. This is a significant advantage over leaking the compressed plaintexts without decompression as the compressed bytes are often not URL-encoded correctly (or at all) by the PDF applications, leading to incomplete or incomprehensible plaintexts. However, due to the inner workings of the deflate algorithms, a complete compressed plaintext can only be prefixed with new segments, but not postfixed. Therefore, a string created using this technique cannot be terminated using a closing bracket, leading to a half-open string. This is not a standard compliant construction, and PDF viewers should not accept it. However, a majority of PDF viewers accept it anyway.

During our security analysis, we identified two standard compliant attack classes which break the confidentiality of encrypted PDF files. Our evaluation shows that among 27 widely-used PDF viewers, all of them are vulnerable to at least one of those attacks, including popular software such as Adobe Acrobat, Foxit Reader, Evince, Okular, Chrome, and Firefox.

You can find the detailed results of our evaluation here.

First, many data formats allow to encrypt only parts of the content (e.g., XML, S/MIME, PDF). This encryption flexibility is difficult to handle and allows an attacker to include their own content, which can lead to exfiltration channels.

Second, when it comes to encryption, AES-CBC – or encryption without integrity protection in general – is still widely supported. Even the latest PDF 2.0 specification released in 2017 still relies on it. This must be fixed in future PDF specifications and any other format encryption standard, without enabling backward compatibility that would re-enable CBC gadgets.

A positive example is JSON Web Encryption standard, which learned from the CBC attacks on XML and does not support any encryption algorithm without integrity protection.