The Mythos model converts browser crashes into working computer exploits with a 72% success rate. Anthropic detailed these capabilities in a 244-page system card. This release challenges the lab's original pledge to avoid leading the frontier. Practitioners must now account for AI-driven vulnerability discovery that bypasses traditional manual exploitation.