Unsealed Emails Show Why Anthropic and the Pentagon Split

A batch of court documents made public this week has put a private argument between Anthropic and the Pentagon on the record, and it explains why one of the most closely watched government AI deals of the year fell apart. The emails, first reported by Gizmodo and The Next Web, show Anthropic chief executive Dario Amodei and a senior defence official trading terms over what the US military could and could not do with Claude.

The official on the other side was Emil Michael, the Pentagon's Under Secretary of Defense for Research and Engineering and a former Uber executive. According to the reporting, Michael wanted access to Claude for what the department called all lawful uses. Amodei held to two limits he had set from the start of the talks: the model was not to be used for fully autonomous weapons, and it was not to be used for domestic surveillance.

Two red lines the Pentagon would not accept

Those two conditions turned out to be the whole fight. Michael told Anthropic the guardrails were, in his words, just not workable, and at one point wrote that there was one more chance to align on core principles before the two sides parted ways. He also pushed back on the idea that a weapons carve-out even made sense, writing that there is no distinction in our world between weapons that are defensive or offensive.

Anthropic did not move. The company has active defence work and has said it will sell to the military, but it draws the line at systems that take a human out of lethal decisions and at surveillance aimed at a domestic population. The emails show those were not talking points bolted on after the fact. They were the reason the negotiation collapsed.

A designation that is hard to square

What gives the documents their edge is the timing. As RealClearDefense reports, Michael emailed Amodei to say the two sides were very close on terms the day after the Pentagon had finalised a supply-chain risk designation against Anthropic. That designation is the sort of label normally reserved for a vendor the government considers a security problem. The judge reviewing the record found it exceedingly difficult to square the friendly private note with the hostile official posture the department was taking at the same moment.

The-Decoder, which also reviewed the exchange, reports that a rival vendor was waiting to pick up the work once talks fell through. That detail matters, because it means the government's leverage was real. Anthropic knew that saying no would cost it the contract, and said no anyway.

Why this one is worth watching

Most fights over military AI happen in the abstract, in white papers and panel discussions about where the lines should sit. This is one of the first times the actual negotiation has been laid out in plain text, with a named company holding a line and a named official trying to move it. The specific claim at issue, that offensive and defensive uses cannot be separated, is exactly the argument safety researchers have warned about for years. Seeing a senior defence official make it in writing is a different thing from reading it in a report.

The episode also sits alongside Washington's wider push to set frontier-AI rules, and it sharpens a question the government has not answered. If a company that already does defence work is treated as a supply-chain risk for refusing autonomous weapons, what signal does that send to every other lab weighing a federal contract? The emails do not settle that. They just make it impossible to ignore.