Anthropic opens its most powerful AI models to more security teams
Anthropic on Tuesday widened a program that lets vetted security teams use its strongest AI models with fewer blocks. The company says partners in its Project Glasswing found at least 129,000 verified software flaws from April to July. More than 33,000 of the flaws it counted are rated critical or high severity.
Anthropic is handing more security teams the keys to its strongest AI. On Tuesday, the company expanded its Cyber Verification Program, which lets vetted defenders use its top models with fewer safety blocks on hacking-related work.
The case for it is a big number. Partners in Project Glasswing, an earlier Anthropic effort to secure critical software, found at least 129,000 verified software flaws between April and July, Reuters reported. Anthropic's own scans of open-source code found 5,500 more. More than 33,000 are rated critical or high severity.
Anthropic says the true count is likely at least five times higher. But the figures come from just 33 partner reports, SiliconANGLE noted, and fewer than half of those partners said how many flaws were patched.
The new program has three levels. All include Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1, plus future models.
The first level covers defense, like responding to attacks and studying malware. Company, school and government security teams can apply, as can researchers with a record of reporting flaws. SiliconANGLE said even a regional hospital could qualify, and Anthropic aims to answer within a few days.
The second level adds authorized break-in testing of systems a group has permission to probe. Only organizations can apply, and reviews take a few weeks. Even here, Anthropic says its filters step in if work drifts toward something like deploying ransomware.
The top level is for a small group cleared to test safety-critical systems, such as power grids, flight systems and bank transfer networks. Anthropic vets each member with the U.S. government. Existing Glasswing members move straight in.
Members must let Anthropic keep their data so it can watch for misuse.
Why this matters: everyday AI models block most hacking tasks because the same skills can attack or defend. When Opus 5.5 launched Sept. 22, most security requests sent to it were routed to an older model, SiliconANGLE reported. Anthropic's April debut of Claude Mythos Preview raised fears that AI could break into software before it was secured, Reuters said.
The program runs through Anthropic's own platform, Google's Vertex AI and Microsoft's Foundry, with narrower access on Amazon's Bedrock.
What we know: Three access levels are open, with the top one vetted alongside the U.S. government.
What we don't: How many of the flaws found have been fixed.
Where outlets agree Reuters and SiliconANGLE both report that Glasswing partners found at least 129,000 verified flaws from April to July, with more than 33,000 rated critical or high.
Where they differ Anthropic says the real impact is at least five times higher, but its count rests on just 33 partner reports and most did not share how many flaws were patched.