Anthropic rejects Pentagon pressure on mass surveillance and autonomous weapons

What happens when the military brings billions of dollars to the table, but demands you strip away all safeguards on autonomous killer robots and mass surveillance?

If you are Anthropic CEO Dario Amodei, the answer is a hard no.

Anthropic published a blunt statement refusing to yield to Department of War demands. The Pentagon insisted contractors accept an "any lawful use" clause and remove core safety guardrails from advanced AI models. In response, Amodei drew two non-negotiable red lines: Claude cannot be used for mass domestic surveillance, and it cannot power fully autonomous weapons that take humans entirely out of the loop.

Threats, labels, and classified networks

Anthropic is hardly anti-military. It was the first frontier AI lab to deploy models on classified U.S. networks and at National Laboratories, with Claude heavily used across defense agencies for intelligence analysis, operational planning, and cyber operations. The company previously even walked away from hundreds of millions in revenue to block firms tied to foreign adversaries.

Yet negotiations with the Pentagon turned ugly. Pentagon CTO Emil Michael pushed to designate Anthropic a "supply chain risk" if it did not cave—a draconian label typically reserved for hostile foreign adversaries. Officials also threatened to use the Defense Production Act to legally force safeguard removals. As Amodei pointed out, the threats are inherently contradictory: one labels Claude a security risk, while the other treats it as too essential to national security to live without.

Defense Secretary Pete Hegseth went further in a Senate hearing, blasting Amodei as an "ideological lunatic." Shortly after, the Pentagon struck classified deployment deals with OpenAI, xAI, Google, Microsoft, Amazon, and Nvidia, while leaving Anthropic out in the cold.

When hallucinations meet the battlefield

Anthropic's stance is grounded in cold technical reality: today's frontier models are simply too unreliable to pull triggers autonomously.

Recent reporting proves how fragile this tech remains. According to CNN, the U.S. military recently scrambled aircraft and had armed teams ready to board a Chinese ship suspected of carrying nuclear components—only to discover moments before the strike that the entire intelligence report was an AI hallucination generated by a chatbot. Bloomberg similarly reported that an investigation into a deadly February strike cited "overreliance on artificial-intelligence technology."

On the surveillance front, existing legal loopholes allow agencies to commercially buy personal tracking and browsing data without warrants. While individual data points look harmless, powerful AI models can stitch them together into an omniscient, automated surveillance apparatus. Anthropic refused to let Claude serve as that engine.

Why it matters

This standoff is the ultimate stress test for AI safety: can a frontier lab's principles survive direct confrontation with state power?

In the short term, Anthropic took a massive financial hit. Defense contracts are among the most lucrative, sticky revenue streams in the enterprise world. Competitors like OpenAI and xAI moved swiftly to capture that territory, even as OpenAI later updated its terms to clarify that its cloud deployment restricts direct autonomous targeting and warrantless spying. By being booted from classified programs, Anthropic surrendered massive runway to its rivals.

We think Anthropic made the right, if painful, strategic move. Fully autonomous kinetic targeting and warrantless domestic spying are toxic third rails. If an unconstrained model causes a catastrophic military blunder or civilian slaughter—a danger highlighted by the recent false-alarm naval raid—the liability and public backlash will be existential. By forcing the Pentagon to publicly bully an American company for refusing to build killer robots, Amodei established an unmatched moat of safety credibility. For everyday developers and enterprise customers, knowing an AI provider won't compromise foundational ethics under extreme government coercion is becoming the ultimate trust signal.