On Tuesday, October 6, 2026, AI technology giant Anthropic announced the expansion of its “Cyber Verification Program” (CVP). The program further integrates the cybersecurity defense project “Project Glasswing” launched in April this year, aiming to allow rigorously vetted cybersecurity professionals and organizations to utilize its most powerful AI models, accelerating software vulnerability detection and reducing potential risks of hacker attacks with minimal interference from blocking classifiers.
The new version of CVP is divided into three access levels based on usage scenarios, each with corresponding review requirements and security controls:
– “Defense Access”: Covering incident response, Security Operations Center (SOC), malware reverse engineering, vulnerability analysis and verification for defense operations. This level is open to a wide range of applicants, including enterprise and government cybersecurity teams, non-profit organizations, universities, critical infrastructure operators, open-source software maintainers, and independent researchers with a track record of public vulnerability disclosures.
– “Red Team Access”: With authorized penetration testing and red team exercise permissions built on defense capabilities. This level is restricted to organizational applicants only, such as internal red teams in enterprises, government cybersecurity units, and professional security and penetration testing companies, excluding individual researchers. Organizations meeting the criteria may temporarily receive “Defense Access” during the review period as a transition.
– “Specialized Access”: This level has the strictest model access restrictions, limited to specific organizations testing highly secure critical systems such as national power grids, aviation navigation systems, telecommunications networks, interbank transfer infrastructure, and governmental administrative networks. Applicants must undergo the most stringent review jointly conducted by Anthropic and the U.S. government, with members of the original “Project Glasswing” directly transitioning to this tier.
Anthropic stated that users at all three levels can access advanced AI resources including Claude Opus 5.5, Sonnet 5.5, and dedicated security model Claude Mythos 5.1. For verified users at each level, interference from blocking classifiers is reduced accordingly.
Nevertheless, the general public versions still maintain stringent security measures, and even users at the highest authorization level will face immediate interception if attempting actions like deploying ransomware with substantial destructive capabilities.
In April of this year, the release of the Claude Mythos model by Anthropic raised concerns in the market about the potential misuse of advanced AI for automated attacks. In response, Anthropic introduced the “Project Glasswing” limiting model dissemination to a few key technological partners for defensive testing.
In their latest announcement, Anthropic revealed that partners participating in the “Project Glasswing” (aimed at safeguarding the world’s most critical software security) scanned key systems using Claude Mythos from April to July, uncovering over 129,000 verified security vulnerabilities, including 33,000 rated as “severe” or “high-risk”. Additionally, the Anthropic team independently scanned open-source projects between April and October, discovering an additional 5,500 vulnerabilities.
Anthropic noted that since many partners have not fully reported statistical data, the above figures are conservative estimates, with the actual impact of vulnerabilities repaired by defense teams expected to be at least five times greater. Several partners also provided feedback that without the analytical capabilities of Claude Mythos, discovering a similar quantity of deep vulnerabilities manually using traditional methods and tools would likely take months or even years.