Pentagon raises security concerns, curtails Anthropic's Mythos expansion

The fight here is not really about who has the smartest chatbot. It’s about who the U.S. government and big institutions trust to put inside sensitive systems. That’s why Anthropic’s situation looks so strange right now. On one side, the Pentagon is still treating the company as a security problem. On the other, officials are openly talking about one Anthropic model — Mythos — as important enough to matter for national defense. (cnbc.com) ### What actually changed? On May 1, the Defense Department said eight companies will deploy frontier AI on classified networks for “lawful operational use.” The list was concrete: SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, Amazon Web Services, and Oracle. Anthropic was not on it. That exclusion mattered because these are the systems tied to IL6 and IL7 environments — the Pentagon’s high-security cloud levels for classified and top-secret work. (defensescoop.com) ### Why was Anthropic left out? Because the Pentagon has not backed off its earlier judgment. Emil Michael, the department’s CTO, said Anthropic is still considered a “supply chain risk.” In plain English, that means defense contractors working with the military have to certify they are not using Anthropic’s Claude models in that work. The underlying dispute(defensescoop.com)ued the administration in March to challenge the blacklist. (cnbc.com) ### So why is Mythos different? Because Mythos seems to have crossed from “general AI product” into “specific cyber capability.” Michael called it a “separate national security moment” and tied it to finding vulnerabilities and patching networks. That is the awkward part of the story — the government is saying Anthropic the supplier is risky, while also signaling that Anthropic’s cyber mod(cnbc.com)e and more like a strategic tool. (cnbc.com) ### Why does that contradiction matter? Because procurement rules and security rules are now colliding. If a company is blacklisted at the supplier level, but one of its models is viewed as mission-critical, the government has to invent workarounds or carve-outs. That is messy. It also tells every AI company chasing defense money that capability alone is not enough. You need a governance posture the customer can live with — especially when the customer is the Pentagon. (cnbc.com) ### What’s happening on Wall Street? Anthropic is not retreating. It is finalizing a roughly $1.5 billion joint venture with Blackstone, Goldman Sachs, and other financial firms to sell AI tools to private-equity-backed companies. Blackstone, Anthropic, and Hellman & Friedman are each expected to put in about $300 million, while Goldman is set to contribute around $150 million. So even while Washington is tightening, private capital is still leaning in. (money.usnews.com) ### Then why did Goldman restrict Claude in Hong Kong? Because trust is fragmenting by geography and use case. Reuters reported that Goldman removed access to Claude for bankers in Hong Kong through its internal AI platform, while other models like Gemini and ChatGPT remained available. The backdrop was data-security sensit(money.usnews.com) now want much finer control over where a model can be used and by whom. (finance.yahoo.com) ### Is this really about safety guardrails? Partly — but not only. Guardrails were the flashpoint in the Pentagon dispute. The larger issue is institutional control. Governments and banks want frontier models, but they also want predictable legal terms, deployment boundaries, and supply-chain confidence. Anthropic built (finance.yahoo.com) (cnbc.com) ### Bottom line Anthropic is learning the hard version of AI competition. The next moat is not just model quality. It’s whether powerful buyers believe they can govern the model, audit the vendor, and defend the relationship if politics or security rules suddenly shift. (defensescoop.com)