Anthropic Opens Powerful AI Model Access for Security Testing
Anthropic is expanding a program allowing vetted cybersecurity professionals to test its most capable AI models with fewer safeguards. The initiative has already helped identify over 100,000 software vulnerabilities.
AI firm Anthropic is broadening its initiative that permits approved cybersecurity experts to test its most powerful artificial intelligence models under reduced security protections. This updated program, named the Cyber Verification Program (CVP), consolidates previous projects. "The initiative has identified over 100,000 software vulnerabilities this year," the company stated, suggesting the actual impact could be at least five times higher.
The program aims to secure critical software globally. From April to July, participants in the "Project Glasswing" phase discovered 129,000 verified vulnerabilities. Anthropic's own open-source scanning efforts found an additional 5,500 during the same period. Over 33,000 of these identified vulnerabilities have been classified as severe or high-risk.
The new CVP features three tiers: Defense, Red Team, and Special. All tiers can access models like Claude Opus 5.5, Claude Sonnet 5.5, and Claude Mythos 5.1. The Defense tier is for roles such as incident response and malware analysis, open to security firms, critical infrastructure operators, and researchers. The Red Team tier grants authorization for penetration testing and is exclusively for organizations.
The Special tier has the fewest restrictions, accessible to a select group of organizations permitted to test crucial systems, including power grids, air traffic control, and interbank transfer infrastructure. Anthropic and the U.S. government jointly vet participants. Existing members of the "Project Glasswing" program will be integrated into this Special tier.