Safety1d ago
Safety
Small Models Match Mythos Vulnerability Detection
Eight open-weights models, including one with only 3.6 billion parameters, detected the same FreeBSD exploit identified by Anthropic's Mythos. A 5.1B-active model also recovered a 27-year-old OpenBSD bug. This suggests high-end model capabilities for vulnerability research are not exclusive to frontier systems. Practitioners can now use cheap, local models for security audits.