Banks warned on Anthropic AI
U.S. officials convened bank executives to warn about cyber‑risks tied to Anthropic’s new, powerful model, signalling that model releases are being treated as potential systemic threats rather than just product launches. The meetings were led by senior officials including the Treasury secretary and the Federal Reserve chair, and reporting suggests the concern is about models being able to exploit software vulnerabilities at scale. Markets reacted: software and cybersecurity stocks fell as investors digested the risk that frontier models could become an operational liability for financial firms. (nytimes.com) (theguardian.com) (finance.yahoo.com)
On April 7, Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell pulled the chiefs of major United States banks into Treasury headquarters in Washington to talk about one company’s artificial intelligence system, not interest rates, inflation, or a bank failure. (Bloomberg Law) The system was Anthropic’s Claude Mythos Preview, a new model the company has not widely released because it appears unusually good at finding software flaws that humans have missed. (New York Times) A software flaw is a hidden mistake in code, like a building with a door the architect forgot to lock. If a model can spot thousands of those doors faster than a human team, it can help defenders patch them or help attackers walk through them first. (Tech Xplore) Reporting on the bank meeting says officials were worried about models that can exploit vulnerabilities at scale, which means testing huge numbers of systems at machine speed instead of one target at a time. That changes cyber risk from a series of burglaries into something closer to industrial automation. (The Guardian) Banks are the first place regulators would worry because they run old and new software side by side, connect to payment networks, and move trillions of dollars every day. A flaw in one bank is a bank problem; a flaw shared across many vendors can become a financial system problem. (New York Times) That is why this meeting stood out. Washington officials were not treating a model launch like a flashy product demo; they were treating it like infrastructure that could change the threat level for the country’s most important institutions. (bloomberg.com)) Investors heard the same message and sold first. On April 9, United States software shares fell after Anthropic’s update revived fears that frontier models could undercut existing software businesses and turn cybersecurity into a harder, more expensive arms race. (Reuters via Yahoo Finance) The sell-off spread into security names because the old pitch for cybersecurity companies was that more digital threats create more demand for their tools. A model that can uncover weaknesses across major operating systems suggests the attackers’ toolkit may be improving faster than the defenders’ products. (cnbc.com)) Anthropic’s own caution added to the alarm. The company held back broad release of Mythos over concerns about what it could expose, which is unusual in a market that normally rewards shipping first and adding guardrails later. (Reuters via Yahoo Finance) The immediate story is about banks, but the precedent is bigger. Once the Treasury Department and the Federal Reserve start summoning chief executives over a named model, every future frontier release will be judged not just on what it can do for customers, but on what it can do to everyone else’s systems. (The Hill)
