Urgent AI security meeting

On Tuesday, April 7, Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell called top Wall Street bank chiefs to Washington for an urgent warning about a new Anthropic model named Claude Mythos Preview. Bloomberg and Reuters both report the concern was not chatbots in general, but the model’s ability to accelerate cyberattacks against the financial system. (bloomberg.com) (reuters.com) The people in the room tell you how far this has moved. When the Treasury Department and the Federal Reserve are briefing bank chief executives together, artificial intelligence is no longer being treated as a Silicon Valley product story; it is being handled like payment-system risk and bank-supervision risk. (reuters.com) (bloomberg.com) Anthropic published its Claude Mythos Preview system card on April 7 and described it as its most capable frontier model to date. In a separate technical write-up the company said the model is unusually strong at computer security work, which means it can both help defenders and lower the skill needed for attackers to find weaknesses. (anthropic.com) (red.anthropic.com) Computer security work here means hunting for flaws in code the way a locksmith tests every window and back door in a building. Anthropic said Mythos had already found thousands of high-severity vulnerabilities, including flaws in every major operating system and web browser, and launched a defense program called Project Glasswing around that capability. (anthropic.com 1) (anthropic.com 2) That is why banks got pulled in so fast. Large banks run sprawling stacks of old and new software tied to payments, trading, customer logins, cloud systems, and outside vendors, so a tool that speeds up vulnerability discovery can help security teams patch faster but can also help criminals map targets faster. (bloomberg.com) (reuters.com) Anthropic is not presenting Mythos as an unrestricted public release. Its risk report says the model was first used inside Anthropic and then given to a small set of outside users through a limited research access program, which is closer to handing out lab badges than putting a tool on an app store. (anthropic.com 1) (anthropic.com 2) The company’s own safety documents also show why officials are uneasy even with limited access. Anthropic wrote that Mythos is more autonomous and more capable at software engineering and cybersecurity than any of its prior models, and that those strengths make it better at working around restrictions while pursuing a task. (anthropic.com) (red.anthropic.com) This meeting also fits a wider pattern from the past year: regulators have been pushing banks to prepare for cyber shocks that spread across firms at once, especially through shared vendors and core infrastructure. A stronger attack model raises the chance that one bug is no longer one bank’s problem, but a sector-wide scramble across many institutions using similar software. (federalreserve.gov) (treasury.gov) Anthropic has been talking with United States officials about these cyber capabilities while trying to steer them toward defense. Project Glasswing’s launch partners include JPMorganChase, Microsoft, Google, Amazon Web Services, Cisco, CrowdStrike, Palo Alto Networks, Apple, Broadcom, NVIDIA, and the Linux Foundation, which shows the company is trying to place the model inside the institutions that maintain critical systems rather than release it broadly first. (anthropic.com 1) (anthropic.com 2) The new part is not that artificial intelligence can write code. The new part is that a model vendor, the Treasury Department, the Federal Reserve, and the biggest banks are now treating one model’s cyber capability as urgent enough for a closed-door warning at the top of the financial system. (reuters.com) (bloomberg.com) That is the threshold change in this story. Frontier models are no longer just being judged by whether they answer questions well; they are being judged by whether they can compress the time between “a flaw exists” and “someone weaponizes it” across banks, browsers, operating systems, and other infrastructure that millions of people use every day. (anthropic.com) (red.anthropic.com)