The Pentagon Wants to Rewrite the Rules of AI

Recently, a tech story much of the media barely covered exposed a seismic shift in US politics. Defense Secretary Pete Hegseth summoned Dario Amodei, CEO of Anthropic, to the Pentagon to discuss the military use of Anthropic’s Claude. But this was not a contractual discussion about $200 million dollars, nor a minor disagreement about terms of use. Instead, the Defense Department is trying to strongarm an AI company into removing the last vestiges of its ethical safeguards so it can use its model for domestic mass surveillance and to wage war using automated weapons without human oversight.

As Axios reported, the Pentagon’s message to Anthropic was clearly “shit or get-off-the-pot”, threatening to designate the company as a “supply chain risk”, a label normally reserved for suppliers linked to adversarial powers. Translated into real language: expulsion from the defense ecosystem and indirect pressure on any company working with the Pentagon to stop using Claude.

To put this in context, Anthropic signed a $200 million contract with the Department of Defense last summer making Claude the only AI model available on classified US military networks. Claude, through Amazon’s Top Secret Cloud and Palantir’s Artificial Intelligence Platform, was used in the military operation to kidnap Venezuela’s President Nicolás Maduro.

This episode changed the tone of the relationship: when Anthropic executives asked for details about the use of Claude in live-fire operations, the Pentagon interpreted the query as ethical. The response was to double down: if Claude is in the system, it will be used at the Pentagon’s discretion.

The Defense Department, which the Trump administration now calls the Department of War, issued an AI strategy in January, outlining its philosophy: all contracts must incorporate “any lawful use” clauses within 180 days, and suppliers must remove any ethical restrictions. The political message is clear: “responsible” AI will be used to wage war, and the Pentagon will decide what is lawful.

The debate is no longer about technology; this is politics. Anthropic has already agreed to remove virtually all but two of its contractual restrictions: a ban on using its model for mass surveillance of US citizens and a refusal to allow its use in fully autonomous weapons systems. And those are precisely the two red lines that the Pentagon wants to erase.

This is not about some futuristic scenario. We know the US government uses mass surveillance: what changes with frontier models like Claude is the scale and the capacity for inference. Even seemingly clear terms like “illegal surveillance of US citizens” become elastic when broad legal interpretations and mass analysis technology come into play. A system capable of processing millions of communications, cross-referencing financial, biometric, and geolocation databases and extracting behavioral patterns in real time not only identifies suspects: it can anticipate behaviors, map entire social networks, and tag emotional states or political trends.

The second red line is even more troubling. The development of lethal autonomous weapons, that is, systems capable of identifying, selecting, and destroying targets without human oversight, is no longer an academic debate. The Pentagon has been moving in that direction for some time, as the Centre for International Governance Innovation notes when describing how the United States is pushing the planet into an age of autonomous weapons. The semantic shift from “human-in-the-loop” to “human-on-the-loop” and finally “human-out-of-the-loop” is progressive, but clear.

Sadly, there seems little appetite in the mainstream media for a debate on this issue: as Lawfare, a lone voice, points out, it should be Congress that decides on the framework for military use of AI, not the Pentagon and private companies.

What is truly worrying is not only what the White House wants to do, but how it is going about it. There is no debate about whether the military should have unfettered access to models capable of policing the population or making autonomous lethal decisions. There is no specific law that regulates the use of generative models in the selection of military targets. The rules are being worked out secretly in contract negotiations between a “secretary of war” and Silicon Valley executives.

When the normative architecture of technologies that affect the life, death, and privacy of millions of people is decided in closed-door meetings under threat of industrial sanctions, the question ceases to be technical and becomes institutional. Who watches the watchers when the state itself threatens to destroy a company that tries to impose limits?

The situation is exacerbated by the fact that other large technology companies have already shown a greater willingness to relax their safeguards to secure government contracts, a dangerous race to the bottom on boundaries and ethics. Take Elon Musk’s, xAI, which now has access to Pentagon systems and will likely be used for weapons systems and mass surveillance if Anthropic won’t play ball. Having said that, even though Dario Amodei has just expressed his refusal to back down to the Pentagon, his company is dropping its Safety Pledge that it would not develop potentially dangerous technology, a stance that earned Dario Amodei a Time magazine cover.

Is the United States still a democracy? Formally, yes. Substantively, it looks as though we are entering what Bertolt Brecht called dark times as the world lurched toward war in the 1930s. A democracy is not defined simply by periodic elections, but by the existence of effective limits to the coercive power of the State. If that power aspires to integrate mass predictive surveillance and automated violence under opaque supervision, the distance between democracy and the police state is dangerously narrowed.

The question is not whether countries should use AI as part of their legitimate right to self-defense. The question is who decides the limits of that use. If the answer is the military apparatus itself, without effective external oversight, then the problem is not about technology, it’s about the rule of law.

UPDATE (02/27/2026): Dario Amodei has answered. And he said NO.

(En español, aquí)

—

This post was previously published on MEDIUM.COM.

—