Anthropic Expands Cyber Program to Test Its Strongest AI

Anthropic has launched a new Cyber Verification Program. It lets vetted cybersecurity pros test its strongest AI models with fewer safety limits, aiming to keep critical software secure worldwide.

Project Glasswing previously helped partners find at least 129,000 verified bugs between April and July. Anthropic's own scans found 5,500 more. Over 33,000 were rated critical or high severity, and the company believes the real impact is at least five times larger.

The new program has three tiers. The defense tier covers incident response and malware analysis. The red team tier adds authorized penetration testing. The specialized tier has the fewest limits and covers critical systems like power grids and aviation.

Anthropic will vet each member together with the US government. Glasswing's current members will join the specialized tier. Claude Mythos Preview had earlier raised worries that AI might attack software before it gets security hardening.