Thought Leadership
Through the Cyber Verification Program, vetted teams receive models with fewer safeguards. According to Anthropic, Glasswing discovered 129,000 vulnerabilities.
AI company Anthropic expanded its Cyber Verification Program (CVP) on October 6. Through it, vetted security professionals can use the company’s AI models for security work with reduced safeguards and fewer blocking classifiers. Accessible models include Claude Opus 5.5, Claude Sonnet 5.5, and Claude Mythos 5.1, as well as future models.
Three Access Tiers
The program comprises three tiers for which organizations can apply. Defense Access is intended for defensive tasks such as incident response, malware reverse engineering, and vulnerability analysis and testing. Red Team Access additionally permits authorized penetration testing and red teaming. Specialized Access has the fewest safeguards and is reserved for a small group of vetted organizations authorized to test security systems.
In an evaluation using the CyScenarioBench benchmark, Claude Opus 5.5’s safeguards at the Defense Access tier blocked 46 out of 50 tasks, according to Anthropic. At the Red Team Access tier, no tasks were blocked, and the model completed 34 out of 50—the same number as without any safeguards. Without CVP access, every task was blocked on the initial request.
129,000 Vulnerabilities via Glasswing
Anthropic states that through Project Glasswing, at least 129,000 vulnerabilities in software were discovered and confirmed between April and July 2026. Scanning open-source software added another 5,500 between April and October. More than 33,000 of the confirmed vulnerabilities were classified as critical or high.
“This is likely an undercount, as it is based on survey data from only a subset of Glasswing partners.”
Anthropic
The company expects the actual impact to be at least five times greater. These figures cannot be independently verified.
Few Vulnerabilities Actively Exploited
Patrick Garrity from security firm VulnCheck evaluated 300 vulnerabilities with CVE identifiers discovered by Anthropic or Project Glasswing. Of those, two—or 0.67 percent—were demonstrably exploited in the wild: a SQL injection in Ghost CMS (CVE-2026-26980) and a session forgery vulnerability in Rejetto HTTP File Server (CVE-2026-61500). Garrity classified 39 of the 300 vulnerabilities as critical and 141 as high.
According to a report by Veracode, AI-generated fixes also carry risks. In tests, around 44 percent of AI code generation tasks introduced a security-relevant vulnerability. Across all models, the average success rate for security reviews stood at 56 percent, barely higher than the 55 percent recorded in the initial report.
(Editorial Team)
