Banks Called Over Anthropic
U.S. regulators and bank chiefs were urgently summoned after worries that Anthropic’s newest AI model could raise cyber and operational risks for financial firms. Officials from the Treasury and the Federal Reserve convened Wall Street leaders to assess potential threats and coordinate a response, a sign regulators are treating advanced AI as a system‑level risk to banks rather than just a product issue. (news.bloomberglaw.com) (semafor.com)
Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell pulled Wall Street chiefs into an emergency meeting this week after Anthropic unveiled a new model called Claude Mythos Preview and warned it could expose previously unknown software flaws. The banks called in were systemically important firms, the small group whose failure could shake the wider financial system. (semafor.com) (news.bloomberglaw.com) Anthropic did not broadly release Claude Mythos Preview. In its April 7 system card, the company said the model showed a large jump in capability and would be kept inside a defensive cybersecurity program with a limited set of partners instead of being made generally available. (anthropic.com) That detail explains the panic. A model that can find bugs faster than human security teams can be a better alarm system for defenders, but it can also hand attackers a map of weak doors and broken locks across the same networks. (anthropic.com) (forbes.com) Banks are unusually exposed because they run old and new technology side by side. A large bank can have payment rails, trading systems, cloud software, and vendor tools all connected together, so one exploited flaw can spread like a power outage jumping between substations. (federalreserve.gov) (newyorkfed.org) The Federal Reserve has already been writing about this in plain terms. Its 2025 cybersecurity report says current and emerging threats to finance include malware, supply-chain risk, and other attacks that can hit supervised institutions and the infrastructure around them. (federalreserve.gov) A February 2026 paper from the Federal Reserve Bank of New York pushed the point further. The authors built an index of financial-system cyber vulnerability and found that risk is concentrated in a small set of the largest firms and in major technology providers that many institutions depend on at the same time. (newyorkfed.org) That is why this was treated as a system problem, not a single-bank problem. If one advanced model can uncover weaknesses across common operating systems, cloud tools, or vendor software, then many banks could be exposed at once because they share the same digital plumbing. (semafor.com) (news.bloomberglaw.com) (newyorkfed.org) Anthropic’s own document says Claude Mythos Preview is its most capable frontier model to date and highlights cybersecurity testing as a central part of the release decision. The company’s choice to limit access became, in effect, a warning flare to governments and banks that capability had moved faster than the usual approval and containment playbooks. (anthropic.com) The meeting also shows how regulators are changing their frame. For the past two years, most bank AI debates focused on chatbots, fraud models, and compliance tools, but this week’s response treated frontier artificial intelligence more like critical infrastructure risk, closer to a payments outage or a cyberattack on market pipes. (semafor.com) (federalreserve.gov) What happens next is likely to be less about banning one model than about controlling who can test it, where it can run, and how quickly newly found flaws are patched across banks and vendors. The emergency call itself was the clearest signal: Washington now sees frontier AI failures as something that could move from one lab to the balance sheets of major banks very fast. (news.bloomberglaw.com) (semafor.com)
