Hardening Firefox with Claude Mythos Preview

· ai security web · Source ↗

Mozilla’s agentic harness using Claude Mythos Preview found and fixed 271 security bugs in Firefox 150, including sandbox escapes, 15- and 20-year-old latent bugs, and race conditions over IPC.

What Matters

  • Agentic harness runs reproducible PoC test cases dynamically, eliminating the false-positive flood that made earlier GPT-4/Sonnet 3.5 static audits unscalable.
  • Sample bugs include a 15-year-old <legend> element UAF, a 20-year-old XSLT hash-table-rehash use-after-free, and a rowspan=0 16-bit bitfield overflow undetected by fuzzers for years.
  • 271 Claude Mythos bugs map to three CVE rollups totaling 316 sub-bugs; Mozilla’s sec-high threshold requires exploitability via normal browsing, not just theoretical code paths.
  • Sandbox escapes dominate the sample — the model was permitted to patch Firefox source restricted to the sandboxed process to craft escape PoCs.
  • Pipeline: parallelized across ephemeral VMs per target file, results written to a bucket, then deduplicated, triaged, and integrated into existing security bug lifecycle.
  • Prototype-pollution sandbox escapes were fully blocked by a prior architectural freeze of privileged-process prototypes — the harness logs confirmed many failed attempts against that hardening.
  • [HN: @tialaramex] Every examined ticket in the public sample touches C++, despite C++ being only ~25% of the Firefox codebase.
  • [HN: @jerrythegerbil] Distinction matters: not all 271 are full exploit-grade vulnerabilities; Mozilla’s “vulnerability” definition is broader than common usage, covering reproducible-but-not-yet-weaponized bugs.

Original | Discuss on HN

Related coverage