Expanding the Cyber Verification Program

Anthropic is consolidating its cybersecurity programs into a single tiered access system for the Cyber Verification Program. The move combines Project Glasswing and the existing CVP into three tiers: Defense Access for standard defensive work like malware analysis and vulnerability validation, Red Team Access for authorized penetration testing and red-teaming, and Specialized Access for testing critical infrastructure like power grids and flight systems. All tiers get access to Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1. The company's reasoning is straightforward—their general models have strict cyber safeguards to prevent misuse, but security teams need more capable tools to actually defend systems. The tiers let them balance this by giving stronger capabilities to vetted organizations while maintaining real-time blocks on genuinely dangerous actions like ransomware deployment or tests on safety-critical systems. The verification requirements scale with access level. Defense Access should take days to approve and covers security teams at companies, nonprofits, universities, government agencies, critical infrastructure operators, smaller security firms, and individual researchers with a track record. Red Team Access takes weeks to review and is limited to organizations, not individuals—it covers authorized penetration testing and red-teaming work. Specialized Access requires deep review in collaboration with the U.S. government and is reserved for organizations testing systems that could affect public safety or markets. Data retention is mandatory for monitoring misuse, though Anthropic plans to offer zero-retention options through a new solution called Enterprise Frontier Safeguards launching later this fall. To validate their approach, Anthropic tested Claude Opus 5.5 against CyScenarioBench, which measures multi-stage cyber operations. Without CVP access, every task was blocked immediately. Defense Access blocked 46 of 50 trials but let four succeed. Red Team and Specialized Access had no blocks, with the model completing 34 of 50 tasks—matching its 67.6% success rate without safeguards. The company claims Project Glasswing partners found at least 129,000 verified vulnerabilities between April and July 2026, with partners saying Claude Mythos accelerated their work by months or years. They expect the actual impact to be five times higher based on incomplete data.