Anthropic has officially expanded its Cyber Verification Program (CVP), opening previously closed testing environments to a broader range of security professionals. This initiative aims to balance the efficiency of AI-assisted security work against potential misuse risks, allowing rigorously vetted organizations and individuals to use Claude's most powerful models with significantly reduced safety filters for tasks such as vulnerability research and malware analysis.
Cross-verification from multiple sources indicates that the new CVP introduces a three-tier access system: Defensive Access for general scenarios, Red Team Access restricted to authorized organizations, and Special Access for high-sensitivity needs. Approved users can leverage advanced cyber capabilities in models including Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1. Disclosed data shows this model has helped identify over 33,000 high or critical vulnerabilities, with some statistics suggesting cumulative discovery of hundreds of thousands of software flaws, drastically shortening security response cycles.
This adjustment marks a strategic shift by AI vendors in cybersecurity from "total lockdown" to "controlled empowerment." By integrating existing programs and establishing appeal channels, Anthropic seeks to maximize the potential of large models in offensive-defensive adversarial contexts while ensuring compliance, providing developers and enterprise security teams with more practically valuable tool support.