Anthropic folds Project Glasswing into a three-tier Cyber Verification Program with fewer blocks for red teams

The program now has Defense Access for security teams, nonprofits, universities and individual researchers with disclosure track records (reviewed in days), Red Team Access for organisations doing authorised penetration testing (reviewed in weeks, individuals excluded), and Specialized Access for a small set of vetted groups testing safety-critical systems such as grids, telecom and flight software, reviewed with the US government. Existing Glasswing members move to the top tier automatically. All tiers cover Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1. Anthropic also published the Glasswing tally: partners found at least 129,000 verified flaws between April and July, plus 5,500 from its own scanning, with the true figure estimated at five times higher. On CyScenarioBench, Red Team Access completed 34 of 50 runs that were fully blocked without program status. Enrollment requires data retention, with zero-retention promised later this fall. For security startups, this is the clearest path yet to building offensive tooling on a frontier model without fighting the classifier, and it sets the bar OpenAI and Google will be measured against.