Anthropic Introduces 3-Tier Cyber Verification Program for AI Access

Anthropic merged its Cyber Verification Program (CVP) and Project Glasswing into a unified offering with three access tiers to its AI models, including Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1. The Defense Access tier supports SOC and incident response teams, malware reverse engineering, and vulnerability validation, available to a broad range of security actors. Red Team Access allows authorized penetration testing with restrictions on harmful activities and is limited to organizations. Specialized Access, vetted with US government collaboration, permits testing of safety-critical infrastructure with minimal cyber blocks. The program enforces data retention for misuse monitoring, with future options for customer-controlled data storage. The CVP is available on Claude Platform, Google Cloud Vertex AI, Microsoft Foundry, and selectively on Amazon Bedrock. The program has identified a significant number of verified vulnerabilities through its participants and scanning efforts.

The program facilitates controlled access to powerful AI models for cybersecurity purposes, enabling defenders and vetted organizations to leverage AI for vulnerability discovery, incident response, and penetration testing. By imposing tiered access and safeguards, Anthropic aims to reduce the risk of AI misuse for harmful cyber activities. The data retention policy supports monitoring and abuse prevention. The initiative potentially accelerates vulnerability identification and remediation efforts but does not introduce a direct vulnerability or exploit itself.

This is not a vulnerability but a security program designed to manage AI access securely. Organizations interested in leveraging Anthropic's AI models for cybersecurity should apply for the appropriate access tier based on their role and needs. Anthropic enforces safeguards and monitoring to prevent misuse, and customers should follow the program's verification and data retention requirements. No direct remediation is necessary as this is a controlled access framework rather than a software flaw.