A few months ago, AI-generated bug reports were the free buffet of security research: cheap to produce, painful to process, and roughly as useful as asking an intern to audit memory allocations on their first day. Mozilla's own engineering post from May 7 describes the before-state with admirable restraint: "It is difficult to overstate how much this dynamic changed for us over a few short months." The same post explains how Claude Mythos Preview and agentic harnesses found and helped fix 271 security vulnerabilities in Firefox 150.
The numbers: 180 sec-high, 80 sec-moderate, 11 sec-low. All from the same pipeline. Part of a total of 423 security bugs fixed across all methods in April 2026 releases.
What changed
Mozilla's Brian Grinstead, Christian Holler, and Frederik Braun published the engineering breakdown. Their explanation for the turnaround centers on two factors.
First, the models got more capable. Claude Mythos Preview is operating at a different level than what was available months ago. AI bug reports went from plausibly-wrong to genuinely reproducible and actionable.
Second, the harness work. Mozilla significantly improved their approach to steering, scaling, and stacking models. The key advances were two: the pipeline can now reproduce real bugs (rather than just suggest them), and it can auto-dismiss false positives.
The first matters for maintainer trust - a report that can't be reproduced wastes time. The second matters for scale - if every 100 AI-generated reports contain 80 false positives, someone has to read 80 reports to find the 20 real ones. Automating that filter changes the economics. It turns out the model was never the bottleneck - the ergonomics around the model were.
Severity definitions
Mozilla's severity system is worth understanding before the numbers land differently than intended. Sec-high means the vulnerability can be triggered with normal user behavior - browsing to a page. Sec-critical is reserved for publicly disclosed or actively exploited issues (none in this batch). Sec-moderate requires unusual steps from the victim.
A sec-high bug is not automatically a practical exploit - 180 of them is not 180 exploits sitting on a shelf ready to ship. Firefox's defense-in-depth architecture means any attacker hoping for RCE has to chain multiple vulnerabilities through several privilege escalation layers - less "click the link, own the machine" and more "PhD thesis submitted in C++". Mozilla classifies sec-high based on crash symptoms detected by AddressSanitizer (use-after-free, out-of-bounds memory) and a threat model that assumes any such bug could be exploited with sufficient effort.
Anthropic's direct contribution
Separately from the Mythos Preview pipeline, Anthropic's Frontier Red team submitted bug reports that resulted in three credited CVEs: CVE-2026-6746, CVE-2026-6757, CVE-2026-6758. These were fixed in other releases and processed through Mozilla's standard external disclosure workflow.
The sample
Mozilla made a calculated decision to unhide a sample of the AI-generated bug reports behind the fixed CVEs - unusual given their normal practice of keeping reports private for months post-fix. Their reasoning: the "extraordinary level of interest" and "urgency of action needed throughout the software ecosystem." The published sample spans multiple browser subsystems and shows the depth and variety of what the pipeline produced.
The question this leaves open
Mozilla's pipeline used Claude Mythos Preview - and here the post conveniently skips the one detail any security team actually needs to know: whether this was a partnership-level arrangement with Anthropic or a standard API tier anyone can pay for. The engineering breakdown is generous with technique and sparse with commercial reality. Congratulations on 271 fixes. Now tell us which tier of the API we need to match them.
Source: Behind the Scenes Hardening Firefox with Claude Mythos Preview, Mozilla Hacks, May 7, 2026. Authors: Brian Grinstead, Christian Holler, Frederik Braun.