Will It Mythos?
Points and comments are a snapshot, not live.
Benchmark finds several public models match Mythos on some bugs, but none find all nine.
The author built a benchmark to test if public models can find security bugs that Mythos claims to be uniquely good at. Nine bugs from Mythos's own documentation were selected, each verified to be identifiable by a top model when given a hint. Models were given the relevant file and full repo access, but no hints about what to look for. Results show Qwen 3.6 27B, DeepSeek, MiMo, and Gemma 4 MoE found 4/9 bugs, matching Opus 4.8 and GPT 5.5 Pro. Mistral Medium, Laguna M.1, and Haiku found none. Mythos itself found four bugs no model in the benchmark found. The author notes the harness is naive and suggests better tooling may improve results.
What commenters are saying
Commenters split into camps: those who believe Mythos (Fable) is genuinely superior for security work, and those who attribute its performance to better persistence or tooling rather than raw intelligence. One commenter reports Fable 'reported like a colleague' and required far less handholding on spatial reasoning tasks. Another found Codex outperformed Fable on optimized Rust code, with Fable acknowledging Codex's improvements 80% of the time. Several note that Opus is 'terrifying at infosec' but warn against assuming these models can write secure systems from scratch. The thread surfaces a specific claim: Fable's primary advantage over Opus seems to be persistence, not global intelligence.