Anthropic is widening access to Claude models with reduced safety restrictions for vetted security work. Its Cyber Verification Program is intended for vulnerability research, malware analysis, incident response and authorized penetration testing—tasks that public models may block because the same techniques can be misused.
The program has three levels. Defense Access is available to corporate security teams, government agencies, universities, critical-infrastructure operators, open-source developers and individual researchers. Red Team Access permits authorized attack simulations and is limited to organizations. Specialized Access covers sensitive systems such as power grids, flight controls and banking infrastructure, with applicants reviewed by Anthropic and the US government.
Anthropic says participants in the predecessor Project Glasswing found at least 129,000 confirmed vulnerabilities from April through July 2026, including more than 33,000 rated high-severity or critical. Those figures came from surveys of a subset of partners and are company-reported. Reduced model safeguards can help defenders investigate realistic threats, but access controls and authorization remain essential because the underlying capabilities are dual-use. The tiered structure is an attempt to match that risk to the sensitivity of the work.