Finance ministers and central bankers gathered at the International Monetary Fund's spring meetings in Washington this week pressed Anthropic and U.S. officials over Claude Mythos, a new artificial-intelligence model the company says can find and exploit security flaws in major operating systems, as the U.S. Treasury urged large banks to test their defenses before any public release.

The intervention marks the first time a cohort of Group of Seven finance chiefs has treated a single commercial AI system as a systemic risk to financial infrastructure. Anthropic has withheld Mythos from general release and instead routed it to tech giants and more than 40 critical-software operators through an initiative it calls Project Glasswing. On Thursday the company shipped Claude Opus 4.7, a less capable successor it says will let it study real-world safeguards before any broader Mythos deployment.

Washington meetings

Canadian Finance Minister Francois-Philippe Champagne told the BBC that Mythos had been discussed extensively at the IMF meetings. "Certainly it is serious enough to warrant the attention of all the finance ministers," he said, adding that the model represented an "unknown, unknown." The U.S. Treasury confirmed it had raised the issue with major banks and encouraged them to stress-test their systems in advance of any public release.

Bank of England Governor Andrew Bailey told the BBC the development had to be taken seriously. "We are having to look very carefully now what this latest AI development could mean for the risk of cyber crime," he said. Barclays Chief Executive C.S. Venkatakrishnan told the broadcaster: "It's serious enough that people have to worry."

What Mythos does

Anthropic revealed Mythos Preview in early April, saying internal red-teamers found it "strikingly capable at computer security tasks." The company said the model had already surfaced thousands of high-severity vulnerabilities, including at least one that had gone undetected for 27 years, across every major operating system and web browser. Anthropic gave 12 technology companies early access through Project Glasswing, including Amazon Web Services, Apple, Microsoft, Google, Nvidia, Broadcom and CrowdStrike, and has extended access to more than 40 organizations that run critical software. CNBC reported Thursday that the Glasswing launch has triggered a series of meetings among Trump administration officials, tech chief executives and bank chiefs over the security implications.

The 4.7 release

Anthropic said Claude Opus 4.7, available Thursday across its Claude products, API and through Microsoft, Google and Amazon cloud services, is its most powerful generally available model but is "less broadly capable" than Mythos Preview on cyber tasks. The company said it trained Opus 4.7 to "differentially reduce" its hacking capabilities and added automated filters that block prompts it deems tied to prohibited or high-risk cybersecurity uses. Pricing matches Claude Opus 4.6, released in February.

"We are releasing Opus 4.7 with safeguards that automatically detect and block requests that indicate prohibited or high-risk cybersecurity uses," Anthropic said in a release. "What we learn from the real-world deployment of these safeguards will help us work towards our eventual goal of a broad release of Mythos-class models."

Skeptics push back

The alarm is not universal. The U.K.'s AI Security Institute, which received preview access, concluded that Mythos could exploit systems with weak security posture but was not dramatically better than Anthropic's prior Claude Opus 4, and said it could not determine whether the model would breach well-defended networks. Ciaran Martin, former head of the U.K.'s National Cyber Security Centre, told the BBC the claims had "really shaken people," while cautioning that AI developers have an incentive to play up capabilities. The Trump administration, which has convened meetings with Anthropic and bank executives, had not publicly commented on the finance ministers' concerns by press time, and Anthropic did not address the IMF discussion in its Thursday release.

The next test is operational. Top bankers are being given Mythos access to probe their own systems, and financial-industry sources cited by the BBC said a rival U.S. AI company could release a comparably powerful model without the same restrictions. Whether Anthropic's staged approach holds may depend on what Opus 4.7's filters catch in the coming weeks.