Small models also found the vulnerabilities that Mythos found

On April 7, Anthropic announced Mythos, a limited-access AI model designed to autonomously find and exploit zero-day vulnerabilities in critical software, backed by 100M USD in usage credits and 4M USD in donations to open-source security organizations. AISLE researchers tested Mythos's showcase vulnerabilities on small, cheap, open-weight models and found they recovered much of the same analysis—including detection of the flagship FreeBSD exploit on a 3.6B-parameter model costing $0.11 per million tokens. The finding challenges the narrative that frontier model scale is required for AI-driven security, arguing instead that the moat lies in system design and security expertise, not the model itself.