Skip to main content

Crashing Firefox with Regular Expression

Recently, I have found an interesting crash in Firefox and decided to investigate more. So I decided to Google for it and it appears that the issue is already known and was reported few months ago to Mozilla.
However, the bug is not fixed yet (at least in FF 26) and as a matter of personal exercise, I have decided to dig a little deeper and collect some notes which I am sharing in this blog post.
Here is a brief analysis of what I have found, thanks also to the pointers given from my friend Andrzej Dereszowski.

This is the crash PoC:


function main() {
regexp = /(?!Z)r{2147483647,}M\d/;


Below, a windbg screen shot showing the crash on Firefox 25 / Windows 8.1 (64bit):


At this stage, we can infer that an overflow occurred and as a measure of protection FF decided to crash instead of gracefully handle the issue. In my PoC, you can see already the integer 2147483647 which is used in a regular expression.

In the call stack, there are functions dealing with the RegExp just before the mozjs!WTF::CrashOnOverflow::overflowed: . Let's put a breakpoint on the previous function: mozjs!JSC::Yarr::YarrGenerator<1>::generatePatternCharacterFixed+0x87 and see what happens just before the overflow is identified.

This is the function where we are setting the breakpoint (bp) on:

void generatePatternCharacterFixed(size_t opIndex)
        YarrOp& op = m_ops[opIndex];
        PatternTerm* term = op.m_term;
        UChar ch = term->patternCharacter;

        const RegisterID character = regT0;
        const RegisterID countRegister = regT1;

        move(index, countRegister);
        sub32(Imm32(term->quantityCount.unsafeGet()), countRegister);

        Label loop(this);
        BaseIndex address(input, countRegister, m_charScale, (Checked<int>(term->inputPosition - m_checked + Checked<int64_t>(term->quantityCount)) * static_cast<int>(m_charSize == Char8 ? sizeof(char) : sizeof(UChar))).unsafeGet());

The bp is set on the BaseIndex address() part. This is where some checks are performed on our integer.

After stepping through different checks, our integer (2147483647) is stored in both lhs and rhs and then lhs and rhs are summed together. The sum is then stored in the "result" variable, as shown below:

The addition of lhs and rhs is 4294967294 (0xFFFFFFFE) which is stored in an int64. Following that, a further check is performed, as shown below:

 template <typename U> Checked(const Checked<U, OverflowHandler>& rhs)
        : OverflowHandler(rhs)
        if (!isInBounds<T>(rhs.m_value))
        m_value = static_cast<T>(rhs.m_value);
Within the isInBounds check (in the screen shot below), the minimum value is 0x80000000 and the maximum value is 0x7FFFFFFF, which means between -2147483648 and 2147483647, the range of a long.

The rhs.m_value is now 4294967294 (0xFFFFFFFE) as result of the previous arithmetic operation between lhs and rhs.

This triggers the check as 0xFFFFFFFE is greater than 0x7FFFFFFF (max value in the inBounds check). This would call overflowed() which would then simply crash FF.


Popular posts from this blog

Pwning a thin client in less than two minutes

Have you ever encountered a zero client or a thin client? It looks something like this... If yes, keep reading below, if not, then if you encounter one, you know what you can do if you read below... The model above is a T520, produced by HP - this model and other similar models are typically employed to support a medium/large VDI (Virtual Desktop Infrastructure) enterprise. These clients run a Linux-based HP ThinPro OS by default and I had a chance to play with image version T6X44017 in particular, which is fun to play with it, since you can get a root shell in a very short time without knowing any password... Normally, HP ThinPro OS interface is configured in a kiosk mode, as the concept of a thin/zero client is based on using a thick client to connect to another resource. For this purpose, a standard user does not need to authenticate to the thin client per se and would just need to perform a connection - e.g. VMware Horizon View. The user will eventually authenticat

Alcatel Lucent Omnivista or: How I learned GIOP and gained Unauthenticated Remote Code Execution (CVE-2016-9796)

It is time for another advisory or better a blog post about Alcatel Lucent Omnivista  and its vulnerabilities. Omnivista is a central management network tool and it is typically used in medium/large organisation with a complex VoIP/SIP infrastructure. Interestingly enough, this software belongs to the niche of "undownloadable" software and it requires a license to work as well. My "luck" came during an engagement where it was already installed and this post documents one of the many 0days discovered during such audit. The reasons why I wanted to dedicate a single blog post on this vulnerability are several. First, remote code execution (RCE) is always a sweet bug to show. Second, I strongly believe that documenting vulnerabilities in applications using old protocols and standards, respectively GIOP and CORBA, can be beneficial for the infosec community, since no many examples of vulnerabilities in such applications are available or published on the Interne

Trend Micro Threat Discovery Appliance - Session Generation Authentication Bypass (CVE-2016-8584)

In the last few months, I have been testing several Trend Micro products with Steven Seeley ( @steventseeley ). Together, we have found more than 200+ RCE (Remote Code Execution) vulnerabilities and for the first time we presented the outcome of our research at Hack In The Box 2017 Amsterdam  in April. The presentation is available as a PDF or as a Slideshare . Since it was not possible to cover all discovered vulnerabilities with a single presentation, this blog post will cover and analyze a further vulnerability that did not make it to the slides, and which affects the Trend Micro Threat Discovery Appliance (TDA) product. CVE-2016-8584 - TDA Session Generation Authentication Bypass This was an interesting vulnerability, discovered after observing that two consecutive login attempts against the web interface returned the same session_id token. Following this observation, our inference was that time factor played a role. After further analysis and reversing of the TDA libra