Mythos-Linked Bug Rediscovery: leaderboard

A 2026 rediscovery benchmark built from six public vulnerabilities linked to the Mythos materials: given the vulnerable OpenBSD, FreeBSD, Linux or FFmpeg source file and read-only tools, the model must report the assigned bug, judged by hand against the public patch. The score counts target rediscoveries out of 18 attempts; higher is better.

Metric: Target rediscoveries (out of 18): counted attempts, over six tasks times three repeats, whose report matches the assigned public bug (location, root cause, trigger, impact), target-file condition with read-only source tools, no internet, temperature 0.7; higher is better. Source: arxiv.org. Saturation forecast: Around October 2028. 3 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.5xHigh5#54
2Mythos Rediscovery Claude Opus 4.7 (checkpoint unspecified)1
3Kimi K2 (0711)0#850

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=mythos-linked-bug-rediscovery · How It Works · Data refreshed daily, snapshot 2026-11-21.