Anthropic raises alarm bells

U.S. officials pulled the chiefs of Bank of America, Citigroup, Goldman Sachs, Morgan Stanley, and Wells Fargo into a surprise Washington meeting this week to talk about one artificial intelligence model from Anthropic. Federal Reserve Chair Jerome Powell and Treasury Secretary Scott Bessent were in the room, which is not how software launches usually go. (cnbc.com) The model is called Claude Mythos Preview, and Anthropic says it is unusually good at finding hidden software flaws. Anthropic limited access instead of putting it on the open market because the same skill that helps defenders patch systems can also help attackers break into them. (anthropic.com) (cnbc.com) Anthropic says Mythos can identify and exploit zero-day vulnerabilities, which are software bugs nobody has patched yet, in every major operating system and every major web browser. The company said many of the flaws it found were 10 to 20 years old, and one patched OpenBSD bug was 27 years old. (anthropic.com) That helps explain why bank regulators cared. Big banks run on layers of old code, outside vendors, internal tools, and internet-facing systems, so a model that gets much better at bug hunting starts to look less like a product launch and more like a financial-stability problem. (cnbc.com) (americanbanker.com) Anthropic tried to get ahead of that by creating Project Glasswing, a defense-focused program that gives Mythos to a restricted group instead of the public. Anthropic said Apple, Google, Microsoft, Nvidia, Amazon Web Services, CrowdStrike, and Palo Alto Networks are among the initial partners using it for defensive work. (cnbc.com) The company also briefed U.S. officials before the rollout and said it has been in ongoing talks with the Cybersecurity and Infrastructure Security Agency and the Center for Artificial Intelligence Standards and Innovation. In other words, Anthropic did not just ship a model and wait for Washington to notice. (cnbc.com) At the same time, Anthropic is getting big enough that it is looking beyond software and into chips. The Information, as cited by The Next Web, reported that Anthropic is exploring whether to design its own artificial intelligence chips instead of relying entirely on outside suppliers. (thenextweb.com) That idea makes more sense when you look at the scale. Bloomberg reported on April 6 that Anthropic’s revenue run rate had passed $30 billion, up from $9 billion at the end of 2025, and that more than 1,000 business customers are now spending over $1 million a year. (bloomberg.com) Right now Anthropic is still buying access to other people’s hardware. Bloomberg reported that Broadcom and Google are expanding a partnership that will give Anthropic access to about 3.5 gigawatts of computing power beginning in 2027, using Google tensor processing units, which are Google’s in-house artificial intelligence chips. (bloomberg.com) So the same week Anthropic showed Washington a model powerful enough to trigger a banking-sector warning, it was also being reported as a company considering control over more of its own hardware stack. One story is about offensive and defensive cyber risk today, and the other is about whether Anthropic wants the factories, chips, and capacity to keep pushing that frontier faster tomorrow. (cnbc.com) (thenextweb.com)