Pentagon emails expose the real fight for Anthropic

For months, the standoff between Anthropic and the Pentagon appeared to be a simple dispute over access to the company’s Claude AI model. But newly unsealed court documents reveal a far more consequential battle: who gets to dictate the ethical boundaries of frontier AI within the US military.

The emails, released Tuesday as part of one of Anthropic’s lawsuits against the Department of Defense and first reported by the Wall Street Journal, trace a tense exchange between two key figures. On one side is Anthropic CEO Dario Amodei. On the other is Emil Michael, the undersecretary of defence for research and engineering.

Amodei held a firm line throughout the negotiations. He insisted on specific guardrails governing Pentagon use of Anthropic’s models. Two categories were strictly off-limits: fully autonomous weapons and domestic surveillance. These same boundaries had been central to his earlier public questioning of military applications for his technology.

The Pentagon, however, demanded a far more permissive framework. It pushed for the models to be used for “all lawful uses,” a broad phrase that leaves significant interpretive latitude.

The talks began to deteriorate in January. After weeks of silence from Anthropic, Michael emailed Amodei, stating he was “hoping that we are closer to engaging with your revised POV.” The message was clear: he wanted the company to bend toward the Pentagon’s position. Amodei responded by reiterating his redlines.

Michael’s response was blunt. The proposed guardrails were “just not workable,” he wrote. He gave Anthropic “one more chance to align on core principles” before the two sides would part ways. He also outright rejected Amodei’s distinction between offensive and defensive weapons. “There is no distinction in our world between weapons that are defensive or offensive,” Michael stated, according to emails published by Gizmodo from the court record.

Amodei pushed back against the “all lawful uses” standard. His core objection was that US law explicitly permits domestic surveillance. He told Michael that the Pentagon’s draft language appeared to “completely remove our redlines.”

The very next day, Defence Secretary Pete Hegseth designated Anthropic a supply-chain risk. That label is typically reserved for companies linked to foreign adversaries. The move transformed a months-long policy disagreement into an active legal battle.

Michael’s role in the affair has come under scrutiny. Financial disclosures reveal he held stock in xAI, a direct rival to Anthropic, alongside other AI investments, as reported by The Guardian and ProPublica. Anthropic argues that the blacklisting was an act of retaliation, not a genuine security measure. A federal judge agreed in late March, granting a preliminary injunction and calling the move “classic illegal First Amendment retaliation.” An appeals court reversed that ruling in April, and the case continues. The dispute is just one thread in Silicon Valley’s rapidly deepening entanglement with the military.

The implications extend far beyond Washington. Anthropic is effectively testing whether a private AI company can impose ethical limits on a government customer without losing its contracts. That very question is at the heart of Europe’s own debates over military and surveillance AI. The EU AI Act is wrestling with the same boundaries. The case also feeds the broader sovereignty debate, as European buyers weigh how much control a US lab retains once its models are deployed in national-security work.

For now, Anthropic has steadied itself. It overcame its latest clash with the administration this week, nearing a deal to restore access to a restricted model. The emails, however, lay bare just how close the relationship came to breaking, and why.