Anthropic Expands Cyberdefenders’ Access to Top Models

Anthropic said it is expanding and restructuring its cybersecurity access programs to enable more cyberdefenders to use its most advanced Claude models to secure their systems against potential AI attacks.

Under the changes, announced Tuesday, all members of Anthropic’s Cyber Verification Program will have access to Anthropic’s most advanced Mythos models for some cyber uses, which were previously reserved for members of Anthropic’s Project Glasswing. The Cyber Verification Program gives participants access to the Claude models with reduced safeguards, which would ordinarily block the models from assisting with cyberattacks.

Project Glasswing, which offered access to Anthropic’s most advanced models, is transitioning to become the “specialized access” tier of its cyber program.

The change follows Anthropic’s finding last week that Zhipu AI’s GLM-5.3 model is capable of autonomous cyberattacks, similar to Anthropic’s original Mythos model. Since GLM-5.3 is open-weight, anyone can download it and modify it, including to help with cyberattacks. Anthropic wrote last week in response that “a critical threshold in freely accessible capabilities has now been crossed. GLM-5.3 underscores the urgency of expanding access to advanced frontier models to a broader set of entities to empower cyber defenders.”

Members of the Project Glasswing tier of Anthropic’s program are select organizations that protect critical infrastructure and software. Anthropic approves new members in collaboration with the U.S. government. The government’s “trusted access” program has effectively absorbed Glasswing and OpenAI’s equivalent program, but the criteria the government uses to evaluate new members are opaque.

The lowest tier of the revamped program is “defense access,” which enables groups including open source maintainers, universities and individual researchers to analyze software vulnerabilities and understand malware. The intermediate tier is “red team access,” which also allows organizations to use Claude models for penetration testing, or mock cyberattacks to discover new vulnerabilities.