- Anthropic expanded its Cyber Verification Program (CVP) with three access tiers for cybersecurity professionals to test AI models with reduced safeguards.
- Project Glasswing uncovered at least 129,000 verified software vulnerabilities between April and July 2026, with over 33,000 rated critical or high-severity.
- Only 2 of 300 vulnerabilities discovered by Anthropic have been exploited in the wild, suggesting AI lowers discovery barriers but not all flaws are exploitable.
- Evaluations show CVP tiers block varying task completion rates, with Red Team Access completing 34 of 50 tasks compared to 46 blocked on Defense Access.
- AI-generated patches can introduce new risks, with Veracode reporting that 44% of AI code generation tasks introduced risky vulnerabilities.
On Tuesday, Anthropic announced an expansion of its Cyber Verification Program (CVP), allowing vetted cybersecurity professionals to test advanced AI models with reduced safeguards, as the company reported its Project Glasswing initiative uncovered at least 129,000 verified software vulnerabilities between April and July 2026. The company also found an additional 5,500 verified vulnerabilities between April and October 2026 through open-source scanning efforts.
Anthropic said that of these verified vulnerabilities, more than 33,000 have so far been rated as critical- or high-severity, noting “this is likely an undercount” and the true impact may be at least five times higher. The updated CVP features three access tiers—Defense Access, Red Team Access, and Specialized Access—each granting access to models including Claude Opus 5.5, Claude Sonnet 5.5, and Claude Mythos 5.1.
According to a CyScenarioBench evaluation, safeguards blocked 46 of 50 tasks on Claude Opus 5.5 in the Defense Access tier, while the Red Team Access tier on the same model completed 34 of 50 tasks—the same completion rate as when no safeguards are applied. Anthropic stated, “These evaluations give us confidence that we can make advanced cyber capabilities safely available to a broader set of defenders.”
In an analysis published late last month, VulnCheck researcher Patrick Garrity revealed that only 2 of the 300 vulnerabilities discovered by Anthropic or Project Glasswing have been exploited in the wild. Those two actively exploited flaws are CVE-2026-26980, an SQL injection in Ghost CMS, and CVE-2026-61500, a session forgery in Rejetto HTTP File Server.
Meanwhile, Veracode reported that 44% of AI code generation tasks introduced a risky security vulnerability in tests, with the average security pass rate across models at 56%—barely changed from 55%. 1Password also demonstrated that AI-generated vulnerability patches can themselves introduce new security risks.
✅ Follow BITNEWSBOT on Telegram, Facebook, LinkedIn, X.com, and Google News for instant updates.
