Mozilla uses AI to find thousands of Firefox security bugs

On Thursday, Mozilla published details on how it is using Anthropic’s Mythos model to scan its Firefox browser for software vulnerabilities. The model has discovered thousands of high-severity bugs that required fixing before public release, according to Anthropic. The integration has led to a significant increase in security patching volume. In April 2026, Firefox shipped 423 bug fixes, compared to just 31 exactly a year earlier in April 2025. This shift comes as AI bug-finding tools improve, particularly as agentic systems—AI systems capable of autonomous decision-making—can now assess their own work and filter out false positives.

The model has proven effective at identifying complex sandbox vulnerabilities, which are security flaws in the browser’s isolated execution environment. To find these, the model must write a compromised patch for the browser and then attack the secure software with the new code implemented. Mozilla’s bug bounty program pays up to $20,000 for sandbox vulnerabilities, yet Brian Grinstead, a distinguished engineer at Mozilla, noted that Mythos is finding more sandbox issues than human researchers. Grinstead stated that while Mozilla does receive human reports, they do not arrive at the volume the AI technique can achieve. Mozilla has published details on 12 of the bugs, which include a 15-year-old browser parsing error.

Despite the model’s capabilities, Mozilla is not using AI to fix bugs directly. Instead, the team uses the AI to code patches for human review, serving as a reference rather than being deployed directly. Grinstead emphasized that human oversight remains essential to the process. “For the bugs we’re talking about in this post, every single one is one engineer writing a patch and one engineer reviewing it,” Grinstead said.

The long-term impact of these tools on the broader cybersecurity industry remains uncertain. However, leadership at both companies suggests the technology could alter the balance between attackers and defenders. Anthropic CEO Dario Amodei expressed optimism, stating that if handled correctly, the tools could leave defenders in a better position because there are only so many bugs to find. Grinstead offered a more measured view, noting that while the tool is useful for both sides, having it available shifts the advantage a little bit to defense.

Why it matters

Mozilla’s use of Anthropic’s Mythos model to identify high-severity bugs in Firefox demonstrates a significant shift in AI-assisted cybersecurity, where agentic systems are now capable of filtering results and finding vulnerabilities that previously went undetected for years.