Regulators Alarmed Over Anthropic Model

British and American financial regulators moved to warn banks after Anthropic’s newest artificial intelligence model surfaced serious software flaws in systems the industry relies on. (fstech.co.uk) In Britain, officials from the Bank of England, the Financial Conduct Authority and His Majesty’s Treasury are working with the National Cyber Security Centre on risks tied to Anthropic’s Claude Mythos Preview. People briefed on the talks told the Financial Times that major banks, insurers and exchanges will be warned within the next fortnight. (fstech.co.uk) In Washington, Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell called major bank chiefs into a special meeting on Tuesday, April 7, while they were already in the city for a Financial Services Forum gathering. CNBC reported that Brian Moynihan, Jane Fraser, David Solomon, Ted Pick and Charlie Scharf attended, while Jamie Dimon did not. (cnbc.com) The model at the center of the talks is not a chatbot for consumers. Anthropic says Claude Mythos Preview is a restricted system built for security work, meaning it can read code like an auditor and spot hidden weaknesses before attackers do. (anthropic.com) Anthropic said last week that Mythos Preview had already found thousands of high-severity vulnerabilities, including some in every major operating system and every major web browser. In a separate technical note, the company said the model could identify and exploit zero-day vulnerabilities, which are previously unknown bugs, across those platforms when directed by a user. (anthropic.com) (red.anthropic.com) That has turned an internal technology question into a financial stability question. Banks now use artificial intelligence inside coding assistants, document-processing pipelines, fraud systems and other tools that connect to sensitive infrastructure, according to the British regulatory discussions. (fstech.co.uk) Anthropic has kept the model on a short leash. The company said it rolled Mythos Preview out in limited form through Project Glasswing, a program that includes Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia and Palo Alto Networks, with up to $100 million in usage credits and $4 million in donations to open-source security groups. (anthropic.com 1) (anthropic.com 2) British officials are discussing the issue through the Cross Market Operational Resilience Group, a forum co-chaired by Bank of England executive director Duncan Mackinnon and UK Finance chief executive David Postings. The Bank of England has not activated its faster emergency forum, the Cross Market Business Continuity Group, and is still handling the issue through existing resilience channels. (fstech.co.uk) The scrutiny lands days after the Bank of England and the Prudential Regulation Authority set out their 2026 approach to “safe AI innovation” in finance. It also follows Prudential Regulation Authority meetings in October 2025 where banks were told their monitoring of artificial intelligence models was “not frequent enough,” according to reporting on presentation slides from those sessions. (tlt.com) (fstech.co.uk) Anthropic has framed the same capability as a defense tool and said it briefed senior United States officials before release on the model’s offensive and defensive cyber uses. Regulators are treating that claim as a reason to move faster, not slower, because the same system that helps patch a bank can also show how to break into one. (cnbc.com) (anthropic.com)