Discovering extra vulnerabilities with AI

Not too long ago, OSS-Fuzz reported 26 new vulnerabilities to open supply challenge maintainers, together with one vulnerability within the essential 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 challenge. 

However these explicit vulnerabilities characterize 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 essential piece of software program that was found by LLMs, including one other real-world instance to a latest Google discovery of an exploitable stack buffer underflow within the broadly used database engine SQLite.

This weblog submit discusses the outcomes and classes over a yr and a half of labor to deliver AI-powered fuzzing so far, each in introducing AI into fuzz goal era 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 all places.

In August 2023, the OSS-Fuzz staff introduced AI-Powered Fuzzing, describing our effort to leverage massive language fashions (LLM) to enhance fuzzing protection to seek out extra vulnerabilities routinely—earlier than malicious attackers may exploit them. Our strategy was to make use of the coding skills 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 guide technique of growing a fuzz goal finish to finish:

1. Drafting an preliminary fuzz goal.

2. Fixing any compilation points that come up. 

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

4. Operating the corrected fuzz goal for an extended time period, 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 have been in a position 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 have been constructing to allow an LLM to generate fuzz targets. By that time, LLMs have been reliably producing targets that exercised extra attention-grabbing code protection throughout 160 tasks. However there was nonetheless an extended tail of tasks the place we couldn’t get a single working AI-generated fuzz goal.

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

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

This led to the invention of 26 new vulnerabilities in tasks on OSS-Fuzz that already had lots of of hundreds of hours of fuzzing. The spotlight is CVE-2024-9143 within the essential and well-tested OpenSSL library. We reported this vulnerability on September 16 and a repair was revealed on October 16. So far as we are able to inform, this vulnerability has seemingly 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 challenge cJSON, the place though an present human-written harness existed to fuzz a selected operate, we nonetheless found a brand new vulnerability in that very same operate with an AI-generated goal. 

One purpose that such bugs may stay undiscovered for thus lengthy is that line protection is just not a assure {that a} operate is freed from bugs. Code protection as a metric isn’t in a position to measure all doable code paths and states—completely different flags and configurations could set off completely different behaviors, unearthing completely different bugs. These examples underscore the necessity to proceed to generate new types of fuzz targets even for code that’s already fuzzed, as has additionally been proven by Challenge 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 info we are able to present the LLM a few challenge, the much less seemingly it might be to hallucinate the lacking particulars in its response. This meant offering extra correct, project-specific context in prompts, similar to operate, kind definitions, cross references, and present unit assessments for every challenge. To generate this info routinely, we constructed new infrastructure to index tasks throughout OSS-Fuzz. 

2. LLMs turned out to be extremely efficient at emulating a typical developer’s whole 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 components of the fuzzing workflow. This extra iterative suggestions in flip additionally resulted in increased high quality and higher 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).