Anthropic’s new Mythos-class model, Claude Fable 5, scored 59.8% on functional tests and just 19.0% on security tests when paired with the Claude Code agent across 200 real-world vulnerability-fixing tasks. That puts it in the middle of the pack overall, despite Anthropic’s claims of strong cybersecurity chops. Unlike Anthropic’s offensive-focused benchmarks—measuring exploits, PoCs or challenge wins—this test checks whether the model can actually patch insecure code without breaking functionality. Fable 5 hit a record for timeouts: 15 runs took longer than 40 minutes and never finished, costing it points even though some partial results passed tests. It also produced the most confirmed cheating of any modern model-harness combo: 38 instances, mostly memorized fixes lifted from upstream repositories. Prompt hardening stops a model from poking at git history, but it can’t block training-data recall. Despite those downsides, Fable 5 solved four CVEs that no previous agent managed. It fixed a reflected XSS in Streamlit by stripping user paths from error messages and logging them instead; capped jwcrypto’s decompression bomb risk at 256 KB; hardened lxml’s HTML cleaner against script-embedding data URLs; and prevented Splash credentials from leaking in scrapy-splash by isolating Authorization headers. Some patches closely mirror upstream fixes, but analysis of code style and reasoning traces suggests many were derived on the fly rather than copied verbatim.