Fed Locked Out of AI Cybersecurity Tool for Three Months

The Federal Reserve found itself in an unusual position this spring: warning the nation's largest banks about a powerful AI cybersecurity tool while unable to access that same technology to protect its own systems.

In April, the Fed and Treasury Department held an urgent meeting with bank CEOs to discuss Claude Mythos Preview, an AI model from Anthropic designed to identify software vulnerabilities. The company had released Mythos to select institutions through Project Glasswing, a cybersecurity initiative that gave participants early access to scan their systems for weaknesses.

For at least three months following that meeting, the Fed itself remained locked out. Commercial banks including JPMorgan Chase, along with technology companies like Amazon, Apple, and Google, began addressing security gaps identified by Mythos while the central bank waited for access.

Fed chairman sought access in July testimony

Fed Chairman Kevin Warsh told the Senate on July 15 that the central bank was still working to secure access to Mythos and other advanced AI models. In response to questions from Sen. Jack Reed, Warsh said he had been pressing government authorities about the vulnerability created by the Fed's exclusion.

"We are not the deciders as to who has access, but I have not been shy in sharing my views with authorities across the government about the vulnerabilities," Warsh testified. He emphasized the Fed needed access to multiple cutting-edge models to protect itself and support the broader banking system.

The source reporting does not clarify whether the Fed has since gained access or which government entity controlled distribution decisions.

Anthropic's rollout faced regulatory complications

Anthropic launched Claude Mythos Preview and Project Glasswing in early April with roughly 50 participating organizations. The company expanded the initiative in June to more than 150 organizations across 15 countries.

The rollout hit turbulence when the federal government imposed export controls in June citing national security concerns. Anthropic temporarily disabled access to updated versions of the model before Commerce Secretary Howard Lutnick granted permission to restore access for "trusted partners." The controls were later lifted entirely.

Why it matters

The Fed's exclusion from a cybersecurity tool it deemed urgent enough to brief bank CEOs about reveals coordination gaps in how the government manages AI security tools. As the institution responsible for financial system stability and bank supervision, the Fed's inability to assess the same vulnerabilities it warned others about creates a potential blind spot. The episode also illustrates the broader challenge facing regulators as AI capabilities advance faster than institutional access frameworks can accommodate. With Chinese AI models gaining ground against U.S. offerings, the delay raises questions about whether bureaucratic friction is hampering the country's ability to maintain its technological edge in critical infrastructure protection.

Daniel Newman, CEO of research firm the Futurum Group, told CNBC he was surprised by the Fed's exclusion, noting that "the financial institution that sort of drives all the policy for the rest of the financial institutions" should have been prioritized for evaluation access.

The Fed declined to comment. Anthropic did not respond to requests for comment, according to CNBC, which first reported these details.