Anthropic model sparks finance alarm

U.S. Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell pulled top Wall Street bank chiefs into an urgent Washington meeting this week after officials grew worried that Anthropic’s newest artificial intelligence model could make cyberattacks on banks faster and harder to stop. Reuters and Bloomberg both reported the session, and Reuters said the meeting happened on Tuesday, April 7. (cnbc.com) (bloomberg.com) The model at the center of it is Anthropic’s Claude Mythos Preview, announced on April 7, 2026, as part of a locked-down security effort called Project Glasswing instead of a normal public launch. Anthropic said it kept the model unreleased because its capabilities had jumped far enough that broad access was not safe. (anthropic.com 1) (anthropic.com 2) This is not a chatbot panic about bad homework or fake essays. Anthropic says Mythos can find and exploit software flaws the way a master locksmith can spot a weak hinge from across the room, and it reported high-severity vulnerabilities in every major operating system and web browser. (anthropic.com 1) (anthropic.com 2) Banks run on layers of old and new software stitched together across payment rails, trading systems, mobile apps, cloud services, and vendor networks. A model that can scan code, spot weak points, and suggest working attack paths can compress weeks of hacker labor into hours, which is why finance regulators treated this like infrastructure risk instead of a normal technology product launch. (anthropic.com) (cnbc.com) Anthropic is also telling the story in two directions at once. It says the same model can help defenders patch systems before criminals get there, which is why Project Glasswing includes Amazon Web Services, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, Palo Alto Networks, and the Linux Foundation. (anthropic.com 1) (anthropic.com 2) That partner list explains the tension in Washington. JPMorganChase joined Anthropic’s defensive program, while Reuters reported that JPMorgan chief executive Jamie Dimon did not attend the Treasury meeting, and Bloomberg said chief executives from Citigroup, Morgan Stanley, Bank of America, Wells Fargo, and Goldman Sachs were there. (anthropic.com) (cnbc.com) (bloomberg.com) Anthropic says access is being limited to launch partners and more than 40 additional organizations that build or maintain critical software, with up to $100 million in usage credits and $4 million in donations to open-source security groups. That is a very unusual release shape for a flagship model, and it tells you the company thinks the danger is not hypothetical. (anthropic.com) (anthropic.com) The company had already been briefing U.S. officials before the bank meeting. Reuters reported that Anthropic said it was in ongoing discussions with government officials about the model’s offensive and defensive cyber capabilities, which means the April 7 Treasury session looks less like a surprise reaction and more like a preemptive warning to the sector most exposed to a digital bank run. (cnbc.com) (anthropic.com) Anthropic’s own safety framework helps explain why this landed so hard. Its February 24, 2026 Responsible Scaling Policy says catastrophic risk depends on what the whole industry does, not just one company, so even a cautious lab can still face a collective-action problem if rivals or attackers move faster. (anthropic.com) That is why the most important fact in this story is not that one company built a stronger model. It is that the Treasury Department and the Federal Reserve treated a model release like a financial-stability issue within days, because once software can attack software at machine speed, banks stop looking like office buildings and start looking like exposed code. (cnbc.com) (anthropic.com)