Fed Vice Chair for Supervision Michelle Bowman said Anthropic's Mythos is moving faster than the rules. In remarks at the Financial Stability Oversight Council's AI roundtable, published Friday by the Federal Reserve, she pointed to the AI model as proof that bank supervisors need new playbooks.
She also said the work is already underway, and her speech makes clear it is being driven by what one model from one company can do.
Mythos is an AI model from Anthropic that scans software for cyber weak spots. Bowman said it works at a speed traditional security tools cannot match.
That same speed is what makes it risky. A bank can use Mythos to fix holes in its own systems, and an attacker can use it to find new ones to break in.
Bowman called that two-way use the heart of the supervisors' problem. Banks of all sizes, she said, have asked Fed staff about getting access to the tool.
On April 8, Treasury Secretary Scott Bessent and Fed Chair Jerome Powell met with the country's largest bank CEOs at Treasury headquarters in Washington to talk specifically about Mythos. The closed-door meeting focused on what the model could mean for cyber defenses across the financial system.
Bowman said the conversation is part of a broader push to coordinate across government before the technology gets further ahead of supervisors.
On April 17, the Fed, OCC, and FDIC released SR letter 26-2, which narrows existing model risk management guidance to traditional models and basic AI. Generative and agentic AI - the kind Mythos uses - now sit outside that framework while regulators work on something new.
In English: The old rulebook for how banks build and check their own models will not apply to the latest crop of AI tools. Banks will use other risk-management practices instead.
Bowman chairs the Financial Stability Board's Standing Committee on Supervisory and Regulatory Cooperation. She said the FSB will release a draft for public comment in the third quarter laying out sound practices for AI in banking.
The U.S. Treasury and SEC are working alongside the Fed on that report. Bowman said the aim is to keep U.S. supervision in line with foreign regulators as global banks roll out these tools at the same time.
She also said industry feedback will shape the final version once the draft is public.
Anthropic is privately held, but the publicly traded banks that adopt Mythos will likely face supervisory questions about how they use it, who has access, and what guardrails are in place.
Citigroup (C) CEO Jane Fraser, Morgan Stanley (MS) CEO Ted Pick, Bank of America (BAC) CEO Brian Moynihan, Wells Fargo (WFC) CEO Charlie Scharf, and Goldman Sachs (GS) CEO David Solomon attended the April 8 meeting at Treasury. JPMorgan (JPM) CEO Jamie Dimon was the only major bank CEO who could not attend.
The day after the meeting, Anthropic released a version of Mythos to a select group of companies under an initiative called Project Glasswing. Early participants include Amazon (AMZN), Apple (AAPL), and JPMorgan (JPM).
The open question is whether U.S. banks end up with broader supervised access to Mythos, and on what terms. Bowman said the agencies will keep refining their approach as more capable AI tools come out in the weeks and months ahead. The FSB draft lands in Q3.
