The CEOs of the largest U.S. banks met in Washington last week as usual. They were there for a Financial Services Forum meeting, the kind of routine lobbying gathering they were accustomed to. Dinner, discussions about politics, nothing extraordinary. But then came an unusual request: come to the Treasury Department. Right now.
Waiting for them at the table were Federal Reserve Chair Jerome Powell and Treasury Secretary Scott Bessent. The topic was Anthropic's new artificial intelligence model called Claude Mythos and cybersecurity, which may never be the same because of it.
The CEOs of Citigroup, Morgan Stanley, Bank of America, Wells Fargo, and Goldman Sachs attended. Only one major bank CEO was absent: JPMorgan's Jamie Dimon, who had excused himself from the meeting. But he understood the message discussed there very well. In his annual letter to shareholders, released that same week, he wrote: cybersecurity remains one of the greatest risks, and artificial intelligence will almost certainly make it worse.
Everyone Is Afraid of the Mythos Model
Anthropic, the company behind the Claude chatbot, has built a model capable of something that until now only a handful of human experts could do. Mythos can scan code and search for security flaws in every major operating system and every widely used web browser. And it can not only find them, but actually exploit them.
The test results are almost unbelievable. The model uncovered thousands of previously unknown security vulnerabilities. One of them had been hidden in the OpenBSD system for 27 years. Another had plagued the FFmpeg multimedia library for more than 16 years. None of the developers or security experts had found it during that time. Mythos did—and the cost of the entire operation? Less than $20,000.
Anthropic therefore did something it had never done with any of its products before. It did not release the model at all. Only a closed group of about 40 technology companies has access to it, including Amazon, Apple, Microsoft, Google, Nvidia, and JPMorgan Chase. The initiative was named Project Glasswing, and its goal is to use Mythos's capabilities for defense, not attack.
Leak of Information About Mythos
But the story of Mythos did not begin with a press release. It began with a leak. In late March 2026, internal Anthropic documents were accidentally placed in a publicly accessible data repository. Thousands of files, including a draft blog post describing the new model's capabilities.
The text stated that Mythos was "currently far ahead of any other AI model in cyber capabilities". It said the model foreshadowed a wave of models capable of exploiting vulnerabilities in ways far beyond the abilities of human defenders. That is not something you read every day in an unofficial document from a technology company.
Shares of cybersecurity companies immediately plunged. Anthropic then had to respond, and on April 7, 2026, it officially announced both the model's existence and Project Glasswing. The company will contribute up to $100 million in credits to partners and donate another $4 million directly to organizations working to secure open-source software.
Washington Sounds the Alarm
The Donald Trump administration did not take the situation lightly. White House economic adviser Kevin Hassett publicly confirmed what had initially been reported only by anonymous sources: the meeting did take place, and it was the right decision. "The bankers were in Washington that day for other meetings, and it was a natural thing for Secretary Bessent to do," Hassett told Fox News. "He brought them to the Treasury Department together with Powell and went over the cyber risks with them so that they would be aware of the threats."
The Treasury Department then issued a statement saying that it planned further coordination meetings with other institutions. The Fed declined to comment. Canada's response came almost immediately, with its Finance Department and the Bank of Canada convening their own meeting with bankers. The topic was the same: Anthropic and Mythos.
One Model's Double Life
It is remarkable that Anthropic is both the creator of the threat and a leader in defending against it. The company that built the model and knows what it can do will use it to patch vulnerabilities that hackers might otherwise exploit. And yet it is the very same company that the Pentagon designated a supply-chain security risk and that President Trump said he had thrown out "like dogs."
Anthropic CEO Dario Amodei wrote when announcing Glasswing: "The dangers of getting it wrong are obvious. But if we get it right, we have a real chance to create an internet and a world that are fundamentally safer than the ones we had before the arrival of AI."
JPMorgan Chase, whose CEO did not attend the secret meeting, is in fact one of the key partners in the entire Glasswing project. And the model that frightened Washington is currently helping to patch security holes in the software underpinning the global financial infrastructure.
Sources: cnbc.com and theguardian.com
