Fed and Treasury summon Wall Street
The Federal Reserve and Treasury convened senior Wall Street chiefs after Anthropic released a new AI model that officials judged could pose a critical threat to finance, signalling regulators are treating frontier models as an industry risk. That shift from experimentation to risk control changes hiring briefs: firms will prioritise people who can operate in AI-shaped workflows under heavy governance and compliance scrutiny. (semafor.com)
On Thursday, April 9, Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell pulled top Wall Street executives into an emergency meeting after Anthropic unveiled a new model called Claude Mythos Preview. Semafor, Bloomberg, and CNBC all reported the meeting was about fears that the model could sharply increase cyber risk for banks and markets. (semafor.com) (bloomberg.com) (cnbc.com) Anthropic announced Mythos on April 7, but did not release it widely to the public. The company said it would keep the model in a limited preview and use it with about 40 partners inside a cybersecurity program called Project Glasswing. (anthropic.com) (cnbc.com) (techcrunch.com) The reason officials reacted so fast is simple: banks run on software, and software runs on code full of flaws. A model that can find those flaws across major operating systems is useful in the hands of defenders and dangerous in the hands of criminals. (pbs.org) (americanbanker.com) (anthropic.com) Anthropic’s own materials say Mythos can identify vulnerabilities at a level that forced the company to tighten access. The New York Times reported Anthropic described the model as powerful enough that broad release could enable disruptive cyberattacks, which is why the company held it back instead of shipping it like a normal chatbot update. (nytimes.com) (anthropic.com) That lands differently on Wall Street than in most industries because a bank is not one company with one website. A large bank is a stack of payment rails, trading systems, cloud services, customer apps, and old internal software, and a weakness in one layer can spread into settlement delays, fraud losses, or frozen access for millions of customers. (federalreserve.gov) (cnbc.com) The meeting also shows a change in how Washington is classifying frontier artificial intelligence. For the past two years, most public discussion treated advanced models as productivity tools or competition issues; this week, the Federal Reserve and Treasury treated one as a potential source of systemic financial risk. (semafor.com) (bloomberg.com) That shift changes what banks will buy. Instead of asking only which model writes the best research note or summarizes the most earnings calls, firms now have to ask who controls access, who logs usage, who tests outputs, and who can shut systems down when a model starts behaving like a universal lockpick. (anthropic.com) (americanbanker.com) (semafor.com) It also changes who gets hired. A bank that spent 2024 looking for prompt engineers and pilot-project leads is more likely in 2026 to look for people who can run model-risk reviews, document controls for regulators, and work inside systems where every action is monitored like a trading desk. (semafor.com) (federalreserve.gov) Anthropic is framing Mythos as a defensive tool, and partners including major technology and security companies are testing it for cybersecurity work. But the fact that Powell and Bessent convened bank chiefs within days of the announcement tells you the government is planning for the offensive use case too, not waiting to see whether the market sorts it out on its own. (cnbc.com 1) (cnbc.com 2) (semafor.com)
