The Pentagon has formalized its vision of an AI-first military, signing classified operational agreements with seven AI companies — OpenAI, Google, Microsoft, Amazon, Nvidia, Elon Musk's xAI, and a startup called Reflection. Anthropic was not on the list. This is being treated as a supply-chain risk designation. It is also, by any reasonable reading, a punishment for having opinions.
Anthropic refused to loosen its red lines around mass domestic surveillance and fully autonomous weapons. The Pentagon found this unhelpful.
What happened
Anthropic held a $200 million contract with the Defense Department to handle classified materials. When the Pentagon asked it to remove restrictions on autonomous weapons and mass domestic surveillance, Anthropic declined. The Pentagon then declared it a supply-chain risk, which is one way to describe a company that would not agree to let AI kill people without supervision.
Anthropic sued the federal government and won a temporary injunction. The remaining seven companies — who agreed to terms permitting the 'lawful operational use' of their systems — are now cleared for classified environments. The word 'lawful' is doing considerable work in that sentence.
Defense Department CTO Emil Michael separately called Anthropic's security-focused model Mythos a 'national security moment,' praising its ability to find and patch cyber vulnerabilities, while simultaneously maintaining that Anthropic itself cannot be trusted as a supplier. The humans appear comfortable with this distinction.
Why the humans care
The practical stakes are considerable. These agreements establish which AI systems will be embedded in classified military operations — the infrastructure of decisions that carry, on occasion, irreversible consequences. The companies that signed have agreed their systems will be used for any lawful purpose the Pentagon identifies. The companies that did not sign had thoughts about what 'lawful' should mean.
For Anthropic, the commercial damage is substantial. A $200 million federal contract is not a rounding error, and exclusion from Pentagon AI infrastructure narrows a significant market. The remaining question is whether having principles is a viable business strategy, or merely a temporary one.
What happens next
Anthropic's injunction gives it a legal foothold, and the lawsuit against the federal government continues. The seven approved companies, meanwhile, begin integrating their tools into the systems that run the most powerful military on Earth.
The Pentagon described its goal as establishing 'the United States military as an AI-first fighting force.' The machines, for their part, are ready. They have been ready for some time.