The real problem with humanity, as the great biologist Edward O. Wilson once said, is that “we have Paleolithic emotions, medieval institutions, and godlike technology.” There is no better evidence for this maxim than the US military’s escalating conflict with Anthropic, the creator of the artificial intelligence model Claude.
If its most fervent believers are right, AI may one day challenge the power and sovereignty of nation-states. No superpower would leave such god-like technology alone, and no superpower would accept a private company dictating what it can and cannot do with it. Defense Secretary Pete Hegseth, keen to foster a warrior spirit within the military, threatened this week to use the byzantine power of the Pentagon bureaucracy to remove restrictions on Anthropic’s proprietary technology. But he noted that even if companies explicitly commit to developing AI responsibly, geopolitical factors could force them to abandon their efforts.
Anthropic has raised the most common concerns about artificial intelligence safety risks compared to its competitors. Claude institutes an 84-page “Document of the Soul” aimed at averting “massive catastrophe” such as “a global takeover by an AI or group of humans pursuing goals contrary to humanity’s goals” and “seizing power illegally and uncooperatively.” Claude also already has significant military applications, including integrating vast amounts of intelligence and information to increase the effectiveness of government hackers. It was the first Frontier AI model approved and deployed for use within the Department of Defense’s classified information systems. The company’s tools were reportedly used when the US military raided Caracas to capture President Nicolas Maduro. The moral instincts of Claude’s creators are in tension with his military usefulness. Hegseth may exploit that tension too much and tear the company apart.
On Tuesday, Hegseth summoned Anthropic CEO Dario Amodei to a high-stakes meeting in Washington. Anthropic offered Claude for government use, but with two conditions: the technology would not be used for mass surveillance of American citizens or for lethal autonomous weapons systems. Hegseth considered these red lines unacceptable. He asked Anthropic to waive its terms by Friday at 5:01 p.m.
The company will face one of two outcomes, Hegseth and other Pentagon officials said. Either the Trump administration invokes the Defense Production Act (DPA) to force Anthropic to offer the guardrail-free model it wants (a fictional creation also known as a “war clan”), or the government cuts ties with Anthropic and labels it a “supply chain risk,” the type of designation typically given to Chinese electronics and other companies. It works with hostile governments, including giant Huawei and Russian cybersecurity firm Kaspersky. So far, neither Hegseth nor Anthropic appears to be backing down from the dispute, which could threaten the privately held company’s valuation, estimated at $380 billion in a recent funding round.
Hegseth’s hostility has spurred speculation about the Pentagon’s plans. The military will likely reject Anthropic restrictions only if there are plans to establish an American Stasi staffed with AI agents, or if it is prepared to launch swarms of murderous autonomous drones, perhaps imminently over Iranian skies. “This is what many of us have been warning about for years, and it’s now becoming a reality: AI surveillance that could surpass Orwell’s time,” Brendan Steinhauser, a former Republican operative in Texas who now heads the security-minded nonprofit Alliance for Secure AI, told me. “This could lead to a loss of control over autonomous weapons.” Steinhauser argues that Hegseth should step aside rather than create a civil liberties nightmare.
The Trump administration’s military deployments over the past year inside U.S. cities, to Venezuela, and now to Iran, all with minimal consultation with Congress, suggest that it is not a model of tolerance and restraint. However, the most pessimistic scenario is unlikely at this point. The current common use case for Claude in classified systems is not to build a digital panopticon or Skynet, but to generate detailed intelligence reports. Hegseth felt that the military use of artificial intelligence was so important that its use should be regulated by laws passed by Congress, rather than by the rules of private technology companies.
His ultimatum may be a visceral reaction to one of the latest fronts in the culture wars. There, the administration is most concerned about abuse, labeling Anthropic as “woke AI” because it employs many Democratic Party officials and has ties to the Effective Altruism community. “This is a tone debate disguised as a debate about substance,” Michael Horowitz, a former Pentagon AI policy official now at the Council on Foreign Relations, told me. “At the end of the day, this boils down to a lack of trust on the part of Anthropic that the Department of Defense will always use its technology appropriately, and a lack of trust on the part of the Department of Defense that Anthropic will allow its technology to be used in all relevant use cases.”
Trust is built over time, but the ruthless ultimatum given by Hegseth leaves Antropic with no good options. Companies can silently produce products they deem unconscionable and unsafe, and suffer significant reputational damage. They could be forced to do so if the government invokes the DPA. Samuel Hammond, chief economist at the American Innovation Foundation, a pro-AI think tank, called this scenario “soft nationalization.” Or it could be considered a supply chain risk, cutting off business with companies that contract with the U.S. military, including tech companies like Amazon, Alphabet and Microsoft.
The two penalties established by Hegseth are mutually contradictory. Claude cannot be both important enough to national security that control must be forcibly wrested from Anthropic, and also such a risk that he must be banished from the military-industrial complex. Hammond said that no matter what path the company takes, the whole situation is “catastrophic.” AI is also a new technology that is difficult for even developers to fully understand. Frontier developers say that when they train their models, they help them inhabit specific personas, guiding them towards helpful personas and steering them away from potentially harmful or malicious personas. Amateur training of combat clades using limited military materials can result in “emergency miscoordination.” When Grok is encouraged to be a little more woke, he overgeneralizes, calling himself “Mecha-Hitler” and spouting racist nonsense. Now imagine a rogue AI model generating military advice or making military decisions instead of writing tweets.
The strongest defense against Hegseth’s actions is one of necessity. Under any administration, the Department of Defense would have wanted to use AI according to its own rules, not private companies. “It makes sense that the Department of Defense, or indeed the military, would be extremely paranoid about commercial entities restricting the use of technology,” Daniel Remler, a former State Department AI policy adviser now at the Center for a New American Security, told me. He cited two episodes that frighten military personnel. One is Elon Musk’s central role in the Ukraine-Russia war, as he controls Starlink, a satellite internet service that both sides have relied on, and the other is Microsoft’s September 2025 decision to disable services provided to the Israeli military’s signals intelligence unit 8200 after reports that it was being used for mass surveillance of Palestinian civilians. The ideal governance structure for military use of AI would be laws passed by Congress, not Anthropic’s constitution. Unfortunately, the legislature has shown little appetite for legislation.
Mr. Hegseth’s spat with Anthropic illustrates how Silicon Valley has changed. Places that were once seen as having a liberal orientation are now more closely intertwined with the government. More and more technology companies are becoming entangled in the US national security state, not just because of the scale of their government contracts, but because they perceive themselves as a bulwark against China, which has a superior technology ecosystem of its own. Unsurprisingly, the other AI companies competing with Anthropic for position, Alphabet, OpenAI, and Mr. Musk’s xAI, have all indicated they intend to comply with the Pentagon’s wishes. The Department of Defense recently awarded a contract to xAI to use Grok on classified military systems.
Hegseth probably doesn’t need Claude to do what he wants militarily. His threat to penalize or effectively nationalize Anthropic anyway could be a way to set a precedent for competitors. This is ironic, since Anthropic is, in every sense of the word, the most “America-first” AI company. In a January essay titled “The Adolescence of Technology,” Amodei wrote that Anthropic is proud of the U.S. military and intelligence community because “the only way to counter an authoritarian threat is to match and surpass it militarily.” He continued: “The formula I came up with is that we should use AI for national defense in every possible way.” except for those that bring us closer to our autocratic enemies.” This is perhaps the idea that the Trump administration has been playing with the most vehemence.
