Anthropic's Mythos Fuelled Hundreds of Firefox Fixes and Exposed Decade‑Old Bugs

Mozilla says Anthropic’s Mythos model dramatically improved vulnerability discovery in Firefox, producing hundreds of fixes and detailed disclosures of long-dormant, high-severity bugs;

Mozilla researchers say Anthropic’s Mythos model substantially increased both the quality and quantity of vulnerability reports for Firefox, uncovering hundreds of fixes and several high-severity bugs that had lain dormant for years. Anthropic warned in April 2026 that Mythos could find thousands of serious flaws; Mozilla’s post published Thursday presents concrete results from applying those model — led techniques to a major browser. This change matters because it alters how security teams discover urgent defects and allocate triage resources.

Mozilla disclosed technical details on a subset of findings: 12 publicly detailed bugs include two unusual sandbox — escape vulnerabilities and a 15 — year-old HTML element parsing error. The Firefox team shipped 423 bug fixes in April 2026, up from 31 fixes in the same month a year earlier — a surge Mozilla attributes in part to AI-assisted discovery. Mozilla’s sandbox bounty program pays up to $20,000 for qualifying reports.

Engineers say the shift reflects two developments: the underlying models have become more capable, and researchers have developed prompting and agentic workflows that let models evaluate and filter their own outputs. That combination, Mozilla argues, moved results beyond the low-quality leads and false positives typical of earlier scanners. “These things are actually just suddenly very good,” said Brian Grinstead, distinguished engineer at Mozilla, describing the practical leap in tooling.

Some of the most striking discoveries targeted sandbox isolation, where an effective exploit must chain a crafted code change into a successful escape from a constrained runtime. Demonstrating such a vulnerability requires a model to propose a compromised patch and then use that change in a multi — step exploit attempt — a demanding technical sequence. The ability of agentic systems to chain edits, run tests, and attempt exploitation produced higher — fidelity proof — of-concept reports that demand faster, more serious triage than many past automated leads.

Despite the gains in automated detection, Mozilla has not moved to fully automated patching. Teams ask models to draft candidate fixes, but humans still write and approve deployable patches: for the bugs discussed, one engineer authored each patch and a second engineer reviewed it. That hybrid workflow has increased pressure on engineering and triage resources and raises operational questions about how to integrate model — driven scanning while preserving careful manual review and secure patch development.