Wall Street Told to Brace for Anthropic's Mythos AI

On Tuesday morning, in a meeting pulled together on short notice in Washington, Treasury Secretary Scott Bessent sat down with chief executives from Bank of America, Citi, and Wells Fargo to raise an alarm about Anthropic's newest AI system. Bessent cautioned the group that if the software were deployed inside their internal networks, it could put confidential customer information at serious risk. Three individuals who were briefed on the discussion but not authorized to comment publicly confirmed what took place.

Federal Reserve Chair Jerome H. Powell joined Bessent at the table. Powell has warned publicly in recent weeks about cyberattacks targeting the financial system. Bloomberg reported that the two officials did not just flag dangers, but they also urged the bank leaders to put the model to work scanning for weaknesses in their own infrastructure.

The AI system in question is Claude Mythos Preview, which Anthropic unveiled this week. The company said it is unusually adept at pinpointing software security gaps that skilled human engineers would miss. Anthropic added that the technology is too powerful and too risky for open distribution at this stage. For now, only about 40 organizations, part of a group Anthropic has named "Project Glasswing," have been granted access.

According to the people with knowledge of Tuesday's discussion, attendees learned that the model's proficiency at exposing banking system vulnerabilities could itself become a danger: hostile actors or hackers who obtained those findings could turn them into an attack playbook. Anthropic has said Mythos was never built with cybersecurity as its intended purpose -- the skill surfaced on its own.

Some in the industry have speculated that limiting availability while advertising the model's potency is less about caution and more about generating enterprise demand.

Among the Project Glasswing partners, JPMorgan Chase, the country's biggest bank, was the only one publicly identified at launch. JPMorgan said it planned to use the tool "to evaluate next-generation A.I. tools for defensive cybersecurity across critical infrastructure." Bloomberg's reporting shows that Goldman Sachs, Citigroup, Bank of America, and Morgan Stanley have since begun running their own tests with Mythos.

JPMorgan CEO Jamie Dimon received an invitation to Tuesday's Washington session but was unable to join because of a prior travel commitment, according to a person with direct knowledge.

Speaking to Fox News on Friday, Kevin A. Hassett, Director of the National Economic Council, said officials are moving quickly. "We're taking every step we can to make sure that everybody is safe from these potential risks, including Anthropic agreeing to hold back the public release of the model until our officials have figured everything out," Hassett said. "There's definitely a sense of urgency."

A Treasury spokesperson described the gathering as one Bessent convened "to initiate a process for planning and coordination of our approach to the rapid developments taking place in A.I." The Fed offered no comment. Bloomberg News was first to report that the meeting had occurred.

Logan Graham, an Anthropic executive, released a statement saying the technology would help "secure infrastructure that is critical for global security and economic stability."

Anthropic and the Trump administration are already in a legal fight. The Department of Defense recently classified the company as a "supply chain risk" -- a label that followed collapsed talks over Anthropic's push to restrict the government's use of its AI, especially for military purposes. Anthropic has challenged the classification in federal court, calling it politically driven and without adequate basis.

One arm of the administration has tagged Anthropic as a security threat. Another is directing the nation's top banks toward the company's most capable model.