We Reproduced Anthropic's Mythos Findings with Public Models

https://blog.vidocsecurity.com/blog/we-reproduced-anthropics-mythos-findings-with-public-models

Article

  • Claims to reproduce Anthropic’s Mythos vulnerability-finding results using public models
  • Used GPT and Opus to find same security issues Mythos found
  • Focused on FreeBSD kernel RPC code as test case
  • Prompted models with specific file paths and line ranges

Discussion

  • Top criticism: prompts were highly pointed (file + line range given), not a fair reproduction
  • Commenters compare it to ‘solving Fermat’s Last Theorem after reading the solution’
  • One commenter reproduced findings with deterministic Python static analysis instead
  • Skeptics note Mythos likely found issues without line-number hints

Discuss on HN


Type Link
Added Apr 17, 2026
Modified Apr 17, 2026