Anthropic Expands Cyber Verification Program: After 129,000 Bugs Found, Its Strongest Models Open to Vetted Defenders — With Fewer Safeguards

What Happened

Reuters reported first on October 6: Anthropic is expanding its Cyber Verification Program (CVP), merging two programs it has run in parallel for the past six months — Project Glasswing (which gave organizations securing critical software access to Claude Mythos) and the original CVP (which gave vetted security teams reduced-safeguard access to Opus and Sonnet). Anthropic's official newsroom simultaneously published the announcement (\"Expanding the Cyber Verification Program,\" October 6), confirming that \"qualifying security professionals\" will get \"advanced cyber capabilities and reduced blocking classifiers.\"

When Glasswing launched in April, Anthropic did not release Mythos Preview publicly and opened it only to partner organizations. Six months later: partners found at least 129,000 verified vulnerabilities between April and July, Anthropic's own open-source scanning added 5,500 more between April and October, and over 33,000 were rated critical or high severity. Anthropic says that's an undercount — the data comes from a survey of a limited set of partners, and the true impact is "likely at least five times higher."

Key Facts

1. Three tiers, loosening as you go up: The Defense tier covers incident response and malware analysis: security teams, critical infrastructure operators, open-source maintainers and researchers with a vulnerability-reporting record can apply. The Red Team tier adds authorized penetration testing and red-teaming, but only organizations qualify. The Specialized tier has the fewest restrictions and is reserved for a small group authorized to test safety-critical systems — power grids, flight systems and interbank transfer infrastructure are on the list; Anthropic vets each member together with the US government, and existing Glasswing members move straight into this tier.
2. "Reduced safeguards" means loosened blocking classifiers: All three tiers get Claude Opus 5.5, Sonnet 5.5, Mythos 5.1 and future models, officially described as "reduced blocking classifiers." This is not a new model release but a permission expansion: the same models, in a version with fewer guardrails for security teams.
3. How the numbers are counted: 129,000 (partners, April–July), 5,500 (Anthropic's own open-source scanning, April–October), 33,000+ critical or high severity, "at least 5x" (Anthropic's own estimate, based on a limited-partner survey). The 129,000 are "verified vulnerabilities," not "fixed" ones.
4. The top tier can test power grids and banking systems: The Specialized tier's authorized targets include power grids, flight systems and interbank transfer infrastructure; Anthropic vets each member together with the US government.

Context

Glasswing launched in April with roughly 50 founding partners (AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks and others), expanding in June to about 150 organizations across 15+ countries in power, water, healthcare, communications and hardware. On October 1, SafeBreach announced it had been approved for both Anthropic's CVP and OpenAI's Daybreak Blue (Trusted Access for Cyber) — both labs run "reduced safeguard" channels for security research. Anthropic also launched a Life Sciences Verification Program in September (a vetted, reduced-safeguard regime for biology work).