Small AI models match Anthropic's Mythos in cybersecurity tests
Original: Small models also found the vulnerabilities that Mythos found
AISLE tested Anthropic's showcase Mythos vulnerabilities on small, cheap open-weight models. Eight of eight models detected Mythos's flagship FreeBSD exploit, including a 3.6B parameter model costing $0.11 per million tokens.
Following Anthropic's April 7 announcement of Claude Mythos and Project Glasswing with $100M in credits for cybersecurity, AISLE tested Mythos's flagship vulnerabilities on smaller models. Their tests found that small open-weight models recovered much of Mythos's analysis. Eight models detected the FreeBSD exploit, including a 3.6 billion parameter model. A 5.1B model identified the 27-year-old OpenBSD bug's core chain. On basic security reasoning tasks, small open models outperformed most frontier models from major labs. AISLE, which has been running AI cybersecurity systems since mid-2025, found 15 CVEs in OpenSSL and 5 in curl. The research suggests AI cybersecurity capability is 'jagged' - it doesn't scale smoothly with model size, and no single model dominates across all cybersecurity tasks.
Why This Matters
Demonstrates that advanced AI cybersecurity capabilities may not require expensive, large models
Source
This article summarizes publicly available information from international media. It is not investment advice.