The lab too safe to clear

On May 1, the Department of Defense certified a row of commercial vendors — Amazon, Google, Microsoft, NVIDIA, OpenAI, SpaceX, the startup Reflection AI, and hours later Oracle — to deploy their AI directly on its secret and top-secret networks. Frontier AI is now standing infrastructure on the systems the government guards most closely.

A California judge called the standard for branding an American AI firm a national-security risk 'a pretty low bar.'

The notable name is the one missing. Anthropic, whose Claude was already running on those classified networks through Palantir's defense software, was left out — and not quietly. Months earlier the Pentagon had branded the company a 'supply-chain risk,' a designation built for firms tied to foreign adversaries and never before aimed at an American one. The trigger was Anthropic's refusal to lift two red lines: no fully autonomous weapons without a human deciding to fire, and no mass surveillance of Americans. The safety-conscious lab was labeled a national-security threat for insisting on safety.

Anthropic is fighting it in two federal courts. A California judge called the standard for the designation 'a pretty low bar' and blocked it; a D.C. appeals panel split and let the blacklist stand while the case proceeds. The fight matters beyond one company: it sets the terms on which every frontier lab will negotiate military use, and it shows the government willing to treat a refusal to build weapons as a security risk rather than a virtue.

The same week, staff at Google DeepMind's London office voted to form what is billed as the first union at a frontier AI lab, explicitly to resist military uses of their work. The power AI researchers have long held individually — to refuse a project or walk to a rival lab — is starting to organize into a collective veto.