Senior officials flag Anthropic risk

On April 10, Bloomberg reported that Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell summoned Wall Street leaders to warn that Anthropic’s new model had opened “a new era of cybersecurity” risk. The unusual part was the guest list: top bank executives were being briefed about one company’s artificial intelligence release before most people had even heard the model’s name. (bloomberg.com) The model is Claude Mythos Preview, which Anthropic published on April 7 alongside a system card calling it the company’s “most capable frontier model to date.” Anthropic said it showed a “striking leap” over its previous flagship, Claude Opus 4.6. (anthropic.com) Anthropic’s own red-team unit said Mythos Preview could identify and exploit zero-day vulnerabilities in every major operating system and every major web browser during testing. A zero-day vulnerability is a software flaw the vendor does not know about yet, which makes it more like finding an unlocked side door than picking a lock. (red.anthropic.com) Anthropic also said the model had already found thousands of high-severity vulnerabilities, including some in every major operating system and browser, and used that claim to launch Project Glasswing. Project Glasswing is Anthropic’s effort to put the model in defenders’ hands before similar capabilities spread more widely. (anthropic.com) That is why banks were in the room. Big banks run old and new software at the same time, move trillions of dollars through connected systems, and sit on top of payment rails that attackers target because one weak spot can ripple across many firms. (federalreserve.gov) The warning also appears to have gone beyond Wall Street. CNBC reported that Vice President JD Vance and Bessent held a separate call with Dario Amodei, Sam Altman, Sundar Pichai, Satya Nadella, Elon Musk, George Kurtz, and Nikesh Arora to discuss how to respond if models start scaling in favor of attackers. (cnbc.com) Anthropic has been preparing for this category of risk for months. In June 2025, it said it had activated Artificial Intelligence Safety Level 3 protections for Claude Opus 4, including stronger internal security to make model weights harder to steal. Model weights are the giant bundle of learned parameters that make a model useful, so stealing them is closer to copying the engine than borrowing the car. (anthropic.com) The company has also been arguing in public that language models are moving toward human-level offensive cyber ability in some settings. In July 2025, Anthropic wrote that Claude Opus 4 and Claude Sonnet 4 were being tested at a moment when models were advancing toward that threshold. (red.anthropic.com) What makes this story sharper is that Anthropic’s safety reputation has already taken a hit once this month. On April 1, Bloomberg reported that the company accidentally released internal source code tied to its Claude Code assistant, though Anthropic said no customer data or credentials were exposed. (bloomberg.com) Markets have been reacting to the cyber angle too. Bloomberg reported on March 27 that cybersecurity stocks including CrowdStrike, Palo Alto Networks, and Zscaler fell after concern spread that an Anthropic model in testing might help hackers get around current defenses. (bloomberg.com) So the immediate story is not that Anthropic was hacked, and not that banks were told to stop using artificial intelligence. The story is that senior United States economic officials treated one unreleased or newly released model’s cyber capability as important enough to brief both bank chiefs and top technology executives, which is a level of preemptive concern artificial intelligence companies have rarely triggered before. (bloomberg.com)