Anthropic has announced a restructuring of its cybersecurity initiatives, merging its Project Glasswing and Cyber Verification Program (CVP) into a single, tiered offering. This change, effective as of October 2026, aims to provide more security organizations with access to Anthropic's AI capabilities for system protection.
Project Glasswing and CVP were initially launched in April 2026 alongside the debut of Mythos, Anthropic's frontier AI model. Project Glasswing provided early access to Mythos for partners to identify vulnerabilities within their systems. Anthropic claims that between April and July 2026, its partners identified at least 129,000 verified software vulnerabilities through these programs. Additionally, Anthropic's own open-source scanning efforts reportedly uncovered another 5,500 verified vulnerabilities between April and October.
Of these identified vulnerabilities, Anthropic states that over 33,000 have been rated as critical- or high-severity. The company suggests this is likely an undercount, estimating the true impact to be at least five times higher based on survey data from a subset of Glasswing partners. Despite the identification of 5,674 true positive vulnerabilities, including 3,014 high-severity and 1,522 critical-severity issues, only 516 have been patched according to Anthropic's figures.
The reconfigured program now features three distinct tiers: Defense Access, Red Team Access, and Specialized Access. Anthropic indicates that the purpose of these tiers is to align model capabilities with specific security tasks, with varying levels of restrictions on security-related interactions.
Defense Access is designed for security teams within companies, nonprofits, universities, and government organizations focused on system defense. In this tier, Anthropic's Claude Opus 5.5 model reportedly faced refusals in 46 out of 50 attempts during CyScenarioBench challenges, succeeding only four times.
Red Team Access is tailored for penetration testing and offensive cyber evaluations. Participants in this tier will still encounter model refusals for interactions that could lead to physical harm or widespread disruption. Claude Opus 5.5 completed 34 out of 50 tasks with Red Team Access safeguards enabled, a performance level similar to what is expected from Specialized Access.
Specialized Access is reserved for a limited number of verified organizations authorized to test safety systems that could impact critical infrastructure or markets, such as flight operating systems, power grids, telecom networks, interbank transfer infrastructure, and government administrative networks. This tier offers the fewest model refusals, excluding open-weight models with suppressed guardrails.
For the immediate future, program participants will be required to allow Anthropic to retain their data as part of its AI safety requirements. However, Anthropic's forthcoming Enterprise Frontier Safeguards program is expected to offer zero data retention. Organizations already benefiting from zero data retention with Claude Fable 5.1 or Claude Mythos 5.1 can continue under those terms within the CVP.
