Discovering extra vulnerabilities with AI

Lately, OSS-Fuzz reported 26 new vulnerabilities to open supply undertaking maintainers, together with one vulnerability within the vital OpenSSL library (CVE-2024-9143) that underpins a lot of web infrastructure. The experiences themselves aren’t uncommon—we’ve reported and helped maintainers repair over 11,000 vulnerabilities within the 8 years of the undertaking. 

However these explicit vulnerabilities signify a milestone for automated vulnerability discovering: every was discovered with AI, utilizing AI-generated and enhanced fuzz targets. The OpenSSL CVE is likely one of the first vulnerabilities in a vital piece of software program that was found by LLMs, including one other real-world instance to a current Google discovery of an exploitable stack buffer underflow within the extensively used database engine SQLite.

This weblog submit discusses the outcomes and classes over a yr and a half of labor to carry AI-powered fuzzing thus far, each in introducing AI into fuzz goal technology and increasing this to simulate a developer’s workflow. These efforts proceed our explorations of how AI can remodel vulnerability discovery and strengthen the arsenal of defenders in every single place.

In August 2023, the OSS-Fuzz workforce introduced AI-Powered Fuzzing, describing our effort to leverage giant language fashions (LLM) to enhance fuzzing protection to seek out extra vulnerabilities robotically—earlier than malicious attackers might exploit them. Our strategy was to make use of the coding talents of an LLM to generate extra fuzz targets, that are just like unit assessments that train related performance to seek for vulnerabilities. 

The perfect resolution could be to utterly automate the handbook strategy of creating a fuzz goal finish to finish:

1. Drafting an preliminary fuzz goal.

2. Fixing any compilation points that come up. 

3. Working the fuzz goal to see the way it performs, and fixing any apparent errors inflicting runtime points.

4. Working the corrected fuzz goal for an extended time frame, and triaging any crashes to find out the foundation trigger.

5. Fixing vulnerabilities. 

In August 2023, we lined our efforts to make use of an LLM to deal with the primary two steps. We had been ready to make use of an iterative course of to generate a fuzz goal with a easy immediate together with hardcoded examples and compilation errors. 

In January 2024, we open sourced the framework that we had been constructing to allow an LLM to generate fuzz targets. By that time, LLMs had been reliably producing targets that exercised extra fascinating code protection throughout 160 initiatives. However there was nonetheless an extended tail of initiatives the place we couldn’t get a single working AI-generated fuzz goal.

To handle this, we’ve been bettering the primary two steps, in addition to implementing steps 3 and 4.

We’re now in a position to robotically achieve extra protection in 272 C/C++ initiatives on OSS-Fuzz (up from 160), including 370k+ traces of latest code protection. The highest protection enchancment in a single undertaking was a rise from 77 traces to 5434 traces (a 7000% improve).

This led to the invention of 26 new vulnerabilities in initiatives on OSS-Fuzz that already had lots of of 1000’s of hours of fuzzing. The spotlight is CVE-2024-9143 within the vital and well-tested OpenSSL library. We reported this vulnerability on September 16 and a repair was printed on October 16. So far as we are able to inform, this vulnerability has doubtless been current for twenty years and wouldn’t have been discoverable with present fuzz targets written by people.

One other instance was a bug within the undertaking cJSON, the place although an present human-written harness existed to fuzz a particular perform, we nonetheless found a brand new vulnerability in that very same perform with an AI-generated goal. 

One cause that such bugs might stay undiscovered for therefore lengthy is that line protection just isn’t a assure {that a} perform is freed from bugs. Code protection as a metric isn’t in a position to measure all doable code paths and states—totally different flags and configurations might set off totally different behaviors, unearthing totally different bugs. These examples underscore the necessity to proceed to generate new forms of fuzz targets even for code that’s already fuzzed, as has additionally been proven by Undertaking Zero prior to now (1, 2).

To realize these outcomes, we’ve been specializing in two main enhancements:

1. Mechanically generate extra related context in our prompts. The extra full and related data we are able to present the LLM a couple of undertaking, the much less doubtless it will be to hallucinate the lacking particulars in its response. This meant offering extra correct, project-specific context in prompts, resembling perform, sort definitions, cross references, and present unit assessments for every undertaking. To generate this data robotically, we constructed new infrastructure to index initiatives throughout OSS-Fuzz. 

2. LLMs turned out to be extremely efficient at emulating a typical developer’s total workflow of writing, testing, and iterating on the fuzz goal, in addition to triaging the crashes discovered. Because of this, it was doable to additional automate extra elements of the fuzzing workflow. This extra iterative suggestions in flip additionally resulted in larger high quality and larger variety of right fuzz targets. 

Our LLM can now execute the primary 4 steps of the developer’s course of (with the fifth quickly to come back).