How AI Bug Hunting Works—and Why It Changes Security
AI models can now autonomously scan millions of lines of code and find vulnerabilities that humans and traditional tools missed for decades. Here's how the technology works and what it means for cybersecurity.
The Old Way of Finding Bugs
For decades, cybersecurity researchers have relied on two main methods to find software vulnerabilities: manual code review and fuzzing. Manual review means a human expert reads through source code line by line, looking for mistakes. It is thorough but painfully slow—a skilled auditor might review a few thousand lines per day in a codebase that contains millions.
Fuzzing takes a different approach. Automated tools bombard a program with random, malformed, or unexpected inputs, hoping to trigger crashes that reveal hidden flaws. Google's OSS-Fuzz service, launched in 2016, has used this technique to find over 10,000 security vulnerabilities across more than 1,000 open-source projects. But even OSS-Fuzz achieves only about 30% code coverage on average, leaving vast stretches of code untested.
Both methods share a fundamental limitation: they struggle with logic bugs—flaws that don't cause obvious crashes but allow attackers to manipulate a program's behaviour in subtle, dangerous ways.
How AI Changes the Game
AI-powered bug hunting works differently from traditional tools. Instead of blindly throwing random inputs at software, modern large language models (LLMs) can read and understand code, reason about what it is supposed to do, and form specific hypotheses about where bugs might hide. They then craft targeted test cases designed to confirm or disprove those hypotheses.
This matters because the AI approach combines strengths that were previously separate. Like a human auditor, AI models understand program logic and intent. Like a fuzzer, they can operate at machine speed across enormous codebases. The result is a system that catches bugs traditional tools cannot.
Google demonstrated this hybrid approach by integrating LLMs into OSS-Fuzz. The AI generates fuzz targets—specialised test functions—for under-tested sections of code. In one project, the open-source XML parser tinyxml2, AI-generated tests increased code coverage from 38% to 69% without any manual intervention. Replicating that result by hand would have taken a developer roughly a full working day—and there are thousands of projects to cover.
From Fuzzing to Autonomous Hacking
The latest generation of AI bug hunters goes far beyond smarter fuzzing. Anthropic's Claude Mythos Preview, revealed through the company's Project Glasswing initiative, represents a step-change in capability. The model can autonomously plan multi-step investigations, identify undisclosed vulnerabilities, write exploit code, and chain multiple flaws together into full attack paths—all without human guidance.
The results are striking. Mythos Preview has uncovered thousands of high-severity vulnerabilities across every major operating system and web browser, according to Anthropic. Among the discoveries:
- A 27-year-old bug in OpenBSD, one of the most security-hardened operating systems in existence, exploitable through a simple network connection
- A 16-year-old flaw in FFmpeg, a widely used video processing library, that automated testing tools had triggered five million times without ever catching
- Multiple chained vulnerabilities in the Linux kernel enabling full privilege escalation from ordinary user access to complete system control
On the CyberGym benchmark for vulnerability reproduction, Mythos Preview scored 83.1%, compared to 66.6% for its predecessor model.
The Defender's Dilemma
AI bug hunting creates a paradox that the cybersecurity community calls the dual-use problem. The same capability that helps defenders find and patch vulnerabilities could help attackers discover and exploit them. A model powerful enough to protect every operating system is also powerful enough to compromise them.
Anthropic's response has been to restrict access. Rather than releasing Mythos Preview publicly, the company launched Project Glasswing as a gated defensive programme, sharing the model with over 40 organisations that build or maintain critical software infrastructure. The company has committed up to $100 million in usage credits and $4 million in direct donations to open-source security organisations to support the effort.
The broader industry is moving in a similar direction. Microsoft, Google, and other major technology companies are integrating AI agents into their security workflows—not to replace human analysts, but to handle the sheer volume of code that no team of humans could review manually.
Why It Matters
The software that runs modern infrastructure—operating systems, browsers, encryption libraries, server frameworks—contains billions of lines of code, much of it written decades ago. Traditional methods have proven insufficient to secure it all. AI bug hunting does not eliminate the need for skilled security researchers, but it dramatically expands what is possible: scanning codebases at a scale and depth that were previously unimaginable, and catching flaws that survived undetected for a generation.
The race is now on between defenders deploying these tools to patch vulnerabilities and adversaries seeking to harness the same technology for attack. How that race unfolds will shape the security of the digital infrastructure the world depends on.
