Anthropic Expands Claude Access for Vetted Cyber Teams as Glasswing Finds 129,000 Flaws
Summary: Anthropic expanded its Cyber Verification Program, giving vetted security teams three tiers of access to models including Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1 with reduced safeguards, and said Project Glasswing uncovered at least 129,000 verified vulnerabilities between April and July 2026, plus 5,500 more through open-source scanning. Anthropic said more than 33,000 of those were rated critical or high severity and that the figure is likely an undercount, while VulnCheck's Patrick Garrity found only 2 of 300 sampled vulnerabilities, or 0.67%, had been exploited in the wild.
What's compelling: Anthropic's own numbers put critical or high-severity findings above 33,000, yet the exploitation rate in the sampled set was 0.67% - so the real question for your team is triage capacity, not discovery. Veracode's finding that roughly 44% of AI code generation tasks introduced a risky vulnerability, with a 56% security pass rate, means the same tooling is adding to the queue it is helping you clear.
Read the original at The Hacker News ↗
- Topics
- AI Security
- Mentions
- CVE-2026-26980CVE-2026-61500
- Related