Mozilla quietly ran Claude over Firefox's 10M-line codebase, fixing 400+ security bugs
claudeai · x · 2026-11-19
Claire Vo reveals that before the Fable/Mythos model was released (and pulled), Mozilla quietly tested it against Firefox's 10-million-line codebase — resulting in 400+ security bug fixes, some hiding for over a decade. Distinguished engineer Brian Grinstead breaks down the agentic bug-finding harness: why you can't just point a model at 10M lines, how to design good goal/loop patterns, killing false positives with a verifier, and why "lying to the agent" helps. His verdict: 50% model, 50% setup — and you can build a similar harness in under an afternoon.
More from coding & agent
- Merge's Agent Handler adds credential passthrough: run tools without storing user tokens — shensi · 2026-11-19
- Gemini 4 Neon called best everyday coding model, beating Opus 5.5 — inductionheads · 2026-11-19
- LangChain raises $300M at $3B valuation and open-sources its design agent — BraceSproul · 2026-11-19
- Volcano, from the Kong team, auto-detects production bugs and opens fix PRs — cneuralnetwork · 2026-11-19
- Researcher's AI workflow advice: start small, do key steps manually before offloading — jmschreiber91 · 2026-11-19
- Open-source akyn-sdk turns any data source into an MCP server in 5 minutes — adnan_hashmi · 2026-11-19