Anthropic's Claude Opus AI Model Outperforms Human Teams in Discovering Firefox Vulnerabilities

Mar 9
2 min read

Key Findings

Anthropic's AI model Claude Opus 4.6 discovered 22 security vulnerabilities in the Mozilla Firefox web browser over the course of two weeks.
14 of the 22 vulnerabilities were classified as high-severity, nearly a fifth of all high-severity Firefox issues fixed in 2025.
Mozilla addressed the majority of these vulnerabilities in Firefox 148, released in January 2026.
This demonstrates AI's growing capability to rapidly detect critical security flaws in complex software.

Background

In late 2025, Anthropic evaluated its Claude Opus 4.6 AI model on the Firefox codebase to test its ability to identify complex, high-impact security vulnerabilities. The researchers initially tasked the model with reproducing known historical CVEs from older Firefox versions, which it was able to do successfully.

They then challenged Claude to find new, previously unreported bugs, starting with the Firefox JavaScript engine. Within 20 minutes, the model identified a Use After Free vulnerability, which the Anthropic team validated and reported to Mozilla along with a proposed patch.

Extensive Vulnerability Discovery

As the researchers continued triaging the model's findings, Claude discovered dozens of additional crashes, leading to a total of 112 unique bug reports across nearly 6,000 C++ files in the Firefox codebase.

Anthropic submitted all of their findings to Mozilla, even if they weren't fully confident that each crashing test case represented a security vulnerability.
Mozilla fixed most of the issues, including the high- and moderate-severity vulnerabilities, in Firefox 148, with remaining patches planned for future releases.
The collaboration prompted Mozilla to begin experimenting with AI-assisted security research internally.

Exploit Generation Attempts

To further test Claude Opus 4.6's capabilities, the researchers provided the model with previously reported Firefox vulnerabilities and asked it to generate working exploits.

Claude attempted several hundred exploit generation tests, spending around $4,000 in API credits.
It was only able to successfully produce exploits in two cases, demonstrating that while the model excels at finding vulnerabilities, exploiting them remains a more difficult and costly challenge.
The successful exploits were described as "crude" and would likely be ineffective in real-world scenarios due to security features like sandboxing.

Implications and Importance of AI-Assisted Security

This project highlights the growing potential of AI-powered tools to rapidly identify critical software vulnerabilities. However, it also underscores the need for caution and human verification when relying on AI for security research.

The scale of findings reflects the power of combining rigorous engineering with new analysis tools for continuous improvement.
AI-assisted vulnerability discovery is a powerful new addition to security engineers' toolbox, but it must be accompanied by robust validation and remediation processes.
The successful, albeit limited, exploit generation attempts also underscore the potential risks as AI-enabled offensive capabilities advance, emphasizing the importance of accelerating the find-and-fix process for defenders.

Sources

https://securityaffairs.com/189131/ai/anthropic-claude-opus-ai-model-discovers-22-firefox-bugs.html
https://www.pcmag.com/news/anthropics-claude-finds-more-bugs-in-firefox-than-human-teams
https://x.com/cybernewslive/status/2030694871820812355
https://www.instagram.com/p/DVnNp8EEYp-/