Anthropic uses Claude to find 22 vulnerabilities in Firefox

In a security research partnership with Mozilla, Anthropic identified 22 separate vulnerabilities in the Firefox web browser. Out of these 22 vulnerabilities, 14 of them were classified as high-severity. Most of the identified bugs have already been resolved in Firefox 148, which was released this February, although a few fixes will have to wait for the next release. The joint project, which was published on 2026-03-06, highlights the role of artificial intelligence in auditing secure, open-source software. The partnership shows how developers can collaborate with AI labs to proactively scan codebases.

The Anthropic research team utilized the Claude Opus 4.6 model over a span of two weeks to conduct the security analysis. The team began its work by examining the browser’s JavaScript engine, before expanding the research to other portions of the codebase. According to the Anthropic research team, they focused on Firefox because “it’s both a complex codebase and one of the most well-tested and secure open-source projects in the world.” This allowed the team to evaluate the model’s performance on a secure codebase.

The findings highlighted a significant gap in the model’s capabilities, as Claude Opus was more effective at finding vulnerabilities than creating exploits for them. To test the model’s ability to write functional exploits, the team spent $4,000 in API credits—which are usage units for accessing AI model services—attempting to create proof-of-concept exploits. However, the model succeeded in generating proof-of-concept exploits in only two cases, demonstrating that writing functional exploits remains a difficult task for the AI. This indicates that while the model can identify flaws, it struggles to automate the creation of working exploits.

While the research demonstrates the utility of AI in identifying code defects, it also underscores the practical challenges for open-source maintainers. While AI tools can help secure codebases, they can also bring a flood of bad merge requests—which are proposed changes to a software codebase—alongside the useful ones. This means that while AI can accelerate the discovery of vulnerabilities, it also introduces administrative overhead for the developers managing the codebase.

Why it matters

Anthropic’s experiment demonstrates that large language models can effectively assist in securing complex open-source software, though their ability to generate functional exploits remains limited.