Anthropic says Claude Opus 4.6 found 22 Firefox vulnerabilities in a Mozilla collaboration

What Anthropic reported on X

On March 6, 2026, Anthropic posted that it had worked with Mozilla to test Claude’s ability to discover security flaws in Firefox. The headline claim was substantial: Claude Opus 4.6 found 22 vulnerabilities in two weeks, and 14 of them were classified as high severity. Anthropic added that this accounted for roughly a fifth of all high-severity Firefox bugs Mozilla remediated during 2025.

The linked Anthropic article adds broader context. The company says the work grew from internal model evaluations into a live collaboration with Mozilla. Anthropic reports that its researchers scanned nearly 6,000 C++ files, submitted 112 unique reports, and that most of the relevant fixes have already shipped in Firefox 148, with the remainder scheduled for later releases.

Why this is a meaningful security signal

This is notable because the story is not about a synthetic benchmark or a capture-the-flag contest. It is about a model helping surface real issues in a production browser used by hundreds of millions of people. Browser security has unusually high stakes because browsers routinely process untrusted code and content, so any credible acceleration in vulnerability discovery matters to both vendors and defenders.

Anthropic’s write-up also suggests the workflow is evolving quickly. Mozilla reportedly encouraged the team to submit findings in bulk instead of manually validating every crash first, which implies the volume of model-generated leads was already large enough to force process changes. That is an operational shift, not just a research headline.

What comes next for AI-assisted security research

The immediate implication is that software producers may need new triage pipelines for model-assisted vulnerability reports. If frontier models can generate valid high-severity findings at this rate, security teams will need better tooling to sort, reproduce, prioritize, and patch them quickly. The bottleneck moves from discovery alone to end-to-end remediation.

There is also a strategic caveat. Anthropic’s post argues that models are currently better at finding vulnerabilities than exploiting them, but the same article frames exploit development as an active area of measurement. That means the defensive advantage could narrow over time. For security teams, the practical response is to treat AI-assisted bug hunting as a permanent part of the threat and defense landscape rather than a temporary novelty.

Sources: Anthropic X post, Anthropic Mozilla/Firefox security write-up