Anthropic Launches Claude Access for Security Teams

Anthropic has spent months running a quiet pilot called Project Glasswing, letting select security researchers use Claude without most of the guardrails that normally stop the model from helping with hacking tasks. Now the company is widening that pool. The new Cyber Verification Program opens Claude's less restricted capabilities to a broader set of vetted organizations and individual researchers, covering work like vulnerability research, malware analysis, incident response, and penetration testing.

Anthropic's public Claude models already block most of that work by default, since the same skills that help a defender patch a flaw can help an attacker exploit one. The CVP tries to split the difference by sorting applicants into three access tiers, with the strictest reserved for systems like power grids and banking networks, vetted jointly with the US government. Corporate security teams, universities, government agencies, and solo researchers fall under a lighter tier aimed at defensive work, while a separate tier covers authorized attack simulations for organizations only.

The results from the earlier pilot give some sense of what's at stake once that access widens.