Mozilla used Anthropic's Mythos model to find 271 zero-day vulnerabilities in Firefox over two months, with almost no false positives. That claim matters because previous AI-assisted vulnerability detection was drowning engineers in hallucinated bug reports, forcing human review that erased any efficiency gain.
The breakthrough rested on two things: improvements in Mythos itself, and Mozilla's custom harness built specifically to support the model while it analyzed Firefox source code. Without that harness, the engineers say earlier runs produced what they called 'unwanted slop', plausible-sounding reports that evaporated under scrutiny.
The full post on hacks.mozilla.org is worth reading for the harness architecture details, not just the headline number. The engineering specifics of how Mozilla constrained the model's analysis pipeline are the actual story here, and they have direct implications for any team trying to apply AI to large, real-world codebases.
[READ ORIGINAL →]