Anthropic Opens Claude to Vetted Cyber Teams as Glasswing Flags 129,000 Flaws

Anthropic has announced it is expanding a program that lets vetted cybersecurity professionals work with its advanced models under reduced safeguards and with blocking classifiers switched off. The move follows results from its Project Glasswing initiative: between April and July 2026, the company says, the effort surfaced at least 129,000 verified software vulnerabilities. A further 5,500 or so were found through open-source scanning between April and October.

More than 33,000 of the verified flaws have been rated critical or high severity. Anthropic cautions that this is likely an undercount, drawn from a survey of only some partners, and that the true impact could be at least five times higher.

Three tiers of access

The revamped program, called the Cyber Verification Program (CVP), offers three levels. Defense Access covers defensive work — incident response, malware reverse engineering, and vulnerability analysis and validation. Red Team Access adds authorized penetration testing. Specialized Access carries the fewest safeguards and is reserved for a limited set of organizations cleared to test safety systems themselves. Every tier grants access to Claude Opus 5.5, Sonnet 5.5, Mythos 5.1, and future models.

Under a CyScenarioBench evaluation, Defense Access blocked 46 of 50 tasks on Claude Opus 5.5, while Red Team Access blocked none and completed 34 of 50 — the same rate as with no safeguards at all. Without CVP access, every such task was rejected on the first prompt.

Finding isn't exploiting

Anthropic justifies the release by pointing to the dual-use nature of these tools: the same capabilities an attacker might weaponize can help defenders harden their systems. A recent analysis, however, tempers the picture. VulnCheck researcher Patrick Garrity found that of 300 vulnerabilities discovered by Anthropic and Glasswing, only two — 0.67% — have been exploited in the wild: CVE-2026-26980, an SQL injection flaw in Ghost CMS, and CVE-2026-61500, a session-forgery bug in Rejetto HTTP File Server.

The takeaway: AI is lowering the barrier to vulnerability discovery, but not every flaw it uncovers is exploitable or genuinely damaging. And as 1Password and Veracode have shown, AI-generated patches can introduce fresh risks of their own. Veracode reports that roughly 44% of AI code-generation tasks introduced a risky vulnerability, with the average security pass rate stuck at 56% — barely moved from 55% in its earlier report, even as AI-written code floods pipelines.

Source: The Hacker News