Claude dug up 14 high-severity Firefox bugs in two weeks

If you use Firefox, your browser just quietly dodged a massive security bullet—and the credit goes to an AI.

Anthropic just revealed the results of a joint initiative with Mozilla: Claude Opus 4.6 discovered 22 vulnerabilities in Firefox over just two weeks. Mozilla classified 14 of them as high-severity flaws, representing nearly a fifth of all high-severity Firefox bugs patched throughout 2025. Most of the fixes were shipped to hundreds of millions of users in Firefox 148.0.

Twenty minutes to zero-day

The project started as an extreme stress test. After watching Opus 4.5 nearly max out standard cybersecurity benchmarks in late 2025, Anthropic wanted to point its latest model at a massive, battle-tested codebase. They picked Firefox.

Claude started with the browser’s JavaScript engine, a notoriously sensitive attack surface. In just 20 minutes, the model flagged a critical Use After Free vulnerability—a memory flaw that attackers can use to run malicious code.

By the time human researchers finished validating that first bug, Claude had already generated 50 more unique crashing test cases. Mozilla’s team was impressed enough to ask for bulk submissions without waiting for manual human vetting. In total, Claude scanned nearly 6,000 C++ files and produced 112 unique reports, complete with minimal test cases and candidate patches written by the AI itself.

Great at finding, terrible at exploiting

The obvious nightmare scenario is that malicious actors could use the same setup to generate point-and-click cyberweapons. Anthropic ran an experiment to see if Claude could write working exploits to read and write arbitrary files.

The results should ease some immediate panic. Anthropic burned roughly $4,000 in API credits across hundreds of runs, and Claude only managed to produce functional exploits in two cases.

Those two successes also had a massive asterisk: researchers had to disable modern browser defenses, including the sandbox. Anthropic noted that finding vulnerabilities is currently an order of magnitude cheaper and far more reliable for LLMs than weaponizing them.

Why it matters

We think this partnership provides the first real blueprint for how AI will transform open-source defense.

For everyday users, the takeaway is simple: update to Firefox 148.0 immediately. But for engineering teams, the implications are massive. Open-source maintainers are notoriously understaffed and drowning in code reviews. Having Claude digest 6,000 C++ files in two weeks shows that AI agents can effectively function as an automated audit army.

Anthropic was smart to pair Claude with automated task verifiers—tools that check whether a fix actually patches the bug without breaking existing software. Instead of spamming Mozilla with hallucinated bug reports, they handed maintainers reproducible test cases and working patches.

The looming risk, however, is the offensive asymmetric curve. While weaponizing exploits is still expensive and crude today, a non-zero success rate means the barrier is lowering. Software maintainers won’t have the luxury of ignoring AI code auditors for long; they will have to adopt them just to keep pace with attackers doing the exact same thing.