Anthropic's Mythos AI uncovers decade-old vulnerabilities in Firefox browser

Security researchers at Mozilla have reported that Anthropic's Mythos AI model has identified a significant volume of high-severity vulnerabilities within the Firefox browser. The findings include complex issues such as sandbox vulnerabilities and dormant bugs that have remained in the codebase for over a decade. This development highlights a marked improvement in the capabilities of AI security tools compared to previous generations, which were often plagued by low-quality reports and high rates of false positives.

Following the deployment of Mythos in April 2026, the impact on Firefox's release cycle was immediate and substantial. In that same month, Firefox shipped 423 bug fixes, a stark increase compared to the 31 fixes shipped in the identical period the previous year. Mozilla researchers have published specific details on 12 of the high-severity bugs uncovered, ranging from unusual sandbox vulnerabilities to a 15-year-old error in how the browser parses an HTML element.

The detection of sandbox vulnerabilities is particularly notable given the intricacy required to find them. Historically, identifying these flaws demanded creative, multi-step processes that often yielded very few results despite high financial incentives within Mozilla's bug bounty program, which offers up to $20,000 for such discoveries. However, the latest generation of tools utilising agentic systems can now assess their own work and filter out bad results, allowing Mythos to find these issues at a volume exceeding human researchers' capabilities.

Despite the AI's ability to generate patch code for these issues, the remediation process is not yet fully automatable. Mozilla engineers remain responsible for writing and reviewing all code patches, as the resulting code usually cannot be deployed directly and instead serves as a model for human engineers. Brian Grinstead, a distinguished engineer at Mozilla, confirmed that for the bugs detailed in their report, every single one involved one engineer writing a patch and another reviewing it.

Anthropic CEO Dario Amodei has suggested that widespread adoption of such tools could shift the cybersecurity balance of power in favour of defenders, provided the identified bugs are fixed. He noted that there are only so many bugs to find, implying a better world could exist on the other side of this discovery phase. Conversely, Grinstead offers a more measured view, acknowledging that while the tool is useful for both attackers and defenders, its availability currently shifts the advantage slightly to defence.

The transition from previous AI security tools to this new agentic architecture represents a significant turning point for the industry. While it is still unclear how these emerging capabilities will change the broader balance of power in cybersecurity in the long term, the immediate data suggests a dramatic improvement in the capability of these models to uncover hidden flaws in major software.