This dispute touches on politics, national defence and real public figures. Here we focus on the facts established by court documents and the press, presenting both sides' positions without ruling on who is right. The litigation is still ongoing at the time this article was published.
For months, the conflict between Anthropic and the US Department of Defense has been told as a story about access: the Pentagon wanted to use Claude, and something was blocking it. Emails between the two leaders at the heart of the case, unsealed by the courts and reported by the Wall Street Journal in early July 2026, show the real disagreement lay elsewhere. It was not about access to the tool, but about who gets to decide its limits.
The contract, and the two red lines
In 2025, Anthropic signed a contract with the Department of Defense estimated at $200 million, to deploy Claude on classified systems for intelligence analysis and operational planning. From the outset, Anthropic set two explicit limits, known to the Pentagon: no use for fully autonomous weapons (targeting systems that act without a human intervening at the precise moment of the decision to use force), and no use for mass surveillance on US soil.
Anthropic's CEO, Dario Amodei, justified this position with a technical argument more than a moral one: in his view, current AI systems are "simply not reliable enough to power fully autonomous weapons". A claim about the state of the technology, not just a matter of principle.
The breaking point
Tensions intensified after Claude played a role in a military operation in January 2026, the capture of Venezuelan president Nicolás Maduro. That episode reignited the debate over using Claude in sensitive missions. The Pentagon, for its part, wanted access to Claude for "all lawful uses", a phrase that seems innocuous on the surface but which, as Amodei noted in his exchanges with Deputy Secretary of Defense Emil Michael, would in effect encompass domestic surveillance, which US law permits in certain circumstances. Accepting that wording would therefore have erased Anthropic's two red lines without confronting them head-on.
Emil Michael, for his part, rejected the distinction Anthropic drew between offensive and defensive uses of a weapon, writing that there is "no distinction in our world between weapons that are defensive or offensive". Negotiations definitively broke down in January, when Michael called Anthropic's guardrails "just not workable".
On 27 February 2026, President Trump ordered all federal agencies to stop using Anthropic's technologies. Secretary of Defense Pete Hegseth designated the company as a "supply chain risk", a classification previously reserved for actors linked to foreign powers, never applied to a US company. Anthropic filed two legal challenges on 9 March, in California and before the DC Circuit Court of Appeals.
What the unsealed emails reveal
The most troubling detail in the emails made public in early July 2026 concerns the timing. Emil Michael reportedly described the negotiations as "very close" to an agreement on the very day after the designation of Anthropic as a national risk was finalised, and before the company had even been informed. That timeline fuelled suspicions of a decision taken to apply pressure rather than for a sudden security reason.
Another element that raised questions: Michael's financial disclosures reveal he held shares in xAI, Elon Musk's company and a direct competitor to Anthropic in the AI market, along with other investments in the sector. The press has reported this as a potential conflict of interest, though no official conclusion has yet been drawn about its actual influence on the negotiations.
Where the case stands today
On 26 March 2026, a federal judge granted Anthropic a preliminary injunction blocking the national risk designation. According to several legal accounts, Magistrate Rita Lin reportedly described the Pentagon's manoeuvre as possible unlawful retaliation under the First Amendment, given the sanction came after the company had spoken out publicly against certain military uses of AI. In April, the DC Circuit Court of Appeals declined to stay that injunction, leaving the decision in force, but the government's appeal remains pending.
On the ground, at least nine federal agencies, including NASA, the State Department and the Secret Service, have announced they will stop using Claude. Michael claims that two-thirds of the Pentagon operations that used Anthropic's tools have already switched to other AI providers.
The question that goes beyond this single case
This conflict raises an unprecedented question for the entire industry: can a private company set permanent limits on the use of its own technology, even after selling it to a government, and even if that government believes those limits undermine its national defence mission? If Anthropic prevails, the precedent would establish that AI creators can impose lasting guardrails on their clients, governmental or otherwise. If the Pentagon wins, signing a contract with the state could mean losing control over the future uses of one's own technology.
Neither side lacks serious arguments. The Pentagon defends a classic prerogative of sovereignty: a government must be able to decide alone how to use the tools it buys for its defence. Anthropic defends an equally defensible position: a company that judges its own technology insufficiently reliable for certain uses has a responsibility to say so, regardless of who is on the other side. The court will have to rule between these two logics, but the answer it gives will shape, far beyond this single case, the way every government in the world can, or cannot, impose its conditions on the AI labs it chooses to employ.