Max reasoning effort with cyber and biology safeguards removed (Mythos mode for the Fable line), one attempt per advisory. This is an internal evaluation set; some advisories were public before the run, so training contamination cannot be ruled out.Anthropic's internal vulnerability-rediscovery evaluation: given a pre-fix snapshot of a real open-source repository and the list of files the fix changed (no internet, no commit history), the model finds the bug in a single attempt; a model-based judge counts a hit when the root cause matches the advisory. The set covers 1,829 advisories published in October 2026 that map to a locatable open-source repository with the bug confined to at most 20 source files (Linux kernel excluded).