Urgent Anthropic risk meeting on Wall Street

Scott Bessent and Jerome Powell did not call Wall Street’s biggest banks into Washington to talk about interest rates. They called them in on April 8 to talk about one artificial intelligence model from Anthropic and what it could do to bank networks if the wrong people got it. (reuters.com) The model is called Claude Mythos Preview, and Anthropic released it only to a limited group instead of putting it on the open market. Anthropic said it held back a broad launch because the system could expose previously unknown software flaws that criminals could turn into break-ins. (anthropic.com) Banks worry about cyberattacks for a simple reason: a modern bank is a stack of old and new software tied to payments, trading, customer accounts, and cloud systems. If one weak point gets hit at the wrong moment, the damage can spread the way a power outage jumps from one substation to another. (federalreserve.gov) Anthropic’s own security team said Mythos Preview can identify and exploit zero-day vulnerabilities across every major operating system and every major web browser. A zero-day vulnerability is a hidden software bug that defenders have had zero days to patch because nobody knew it existed yet. (red.anthropic.com) Anthropic also said the model’s cyber strength comes from a broader jump in coding ability, not from a narrow hacking feature bolted on later. A system that can read, edit, and reason through huge codebases can also spot the cracks humans missed. (anthropic.com) That is why this became a Treasury Department and Federal Reserve problem instead of just a Silicon Valley problem. The Treasury Department watches financial stability, and the Federal Reserve supervises the largest US banks, so both agencies care when a new tool could compress months of attack work into hours. (treasury.gov, federalreserve.gov) Reuters reported that Bessent and Powell warned bank chiefs about cyber risks tied to the model, and Bloomberg reported the meeting took place at Treasury headquarters. When the Treasury secretary and the chair of the central bank appear together for an urgent closed-door risk call, the message is that this is being treated less like a product launch and more like a stress event. (reuters.com, bloomberg.com) Anthropic is not pretending the answer is to freeze the model and walk away. On April 7 it launched Project Glasswing with partners including Amazon Web Services, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, Palo Alto Networks, and the Linux Foundation to use the model on defense before attackers catch up. (anthropic.com) Anthropic said it extended access to more than 40 additional organizations that build or maintain critical software infrastructure, and it committed up to $100 million in usage credits plus $4 million in donations to open-source security groups. That tells you the company thinks the safest near-term move is to put the tool in the hands of defenders first and let them patch as fast as possible. (anthropic.com) The uncomfortable part is that Anthropic’s alignment report says Mythos Preview is its best-aligned model yet, while also saying the system is more autonomous and better at working around obstacles than earlier models. In plain English, the model can follow rules better and still be more dangerous because it is more capable. (anthropic.com) So the story here is not that one company built a scary demo. The story is that US financial regulators are now acting as if a frontier model’s behavior can become part of systemic risk, the same category that used to be reserved for bank runs, payment failures, and market plumbing breaking under stress. (reuters.com, federalreserve.gov)