Is Anthropic Really Saving The World From Uncontrolled Military AI?

The conflict between the U.S. Department of War and the artificial intelligence startup Anthropic highlights two major shifts: The growing dependence of governments on private innovation, and the rise of algorithmic weapons whose autonomy could change the very nature of warfare.

Pete Hegseth, the U.S. Secretary of War, has placed Anthropic on the federal “supply chain risk” list. This is an unprecedented decision involving an American company. The two laws cited by the Secretary are meant to prevent sabotage and espionage – acts Anthropic has in no way committed.

The Secretary also banned Pentagon suppliers from working with the startup: “Effective immediately, no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic.” This ban is illegal. The Department can only impose bans related to federal government contracts. Amazon, Google, and Microsoft will therefore continue to sell Anthropic’s models to their customers. Microsoft has gone even further in supporting the company led by Dario Amodei, stating that a judge should issue a temporary restraining order preventing the Pentagon’s order from being enforced.

For its part, Anthropic is suing the Department of War, arguing that the designation is based purely on ideological motives. In fact, Claude, Anthropic’s flagship model, appears to have become indispensable to the U.S. military. It was reportedly used in recent strikes on Iran to analyze intelligence, identify targets, and simulate battle scenarios, even after Donald Trump had banned the Company from all federal activity.

How Dependent Can Governments Become On Private Innovation?

The conflict between the Pentagon and Anthropic began with the Company’s refusal to allow its technology to be used for mass surveillance and to be deployed in autonomous weapons without human oversight. These demands seem reasonable given the enormous risks of fully delegating law enforcement or warfare operations to this technology (see below), especially considering its lack of reliability.

Military decisions should remain in the hands of elected officials, and those officials should not depend on usage conditions set by a supplier, even though it is common for companies to place limits on how governments use their systems. But the way Donald Trump and Pete Hegseth are using their political power is unusual and counterproductive. Why weaken, sometimes out of pure ego, one of America’s leading companies in the most important technology for the future of the world?

If the U.S. government can no longer fully control strategic programs the way it did during the Manhattan Project, the space race, or the creation of ARPANET, it must learn to work with private actors. At the same time, we may be witnessing the rise of a new military-industrial complex built around artificial intelligence and companies such as Anduril and Palantir – one that is far more independent from the state than the one that emerged in the 1950s. In reality, tech companies are becoming geopolitical actors. No company illustrates this better than SpaceX with its Starlink network.

In this context, Dario Amodei has worked hard to persuade a technology community that traditionally leans left in American politics to cooperate with the Department of Defense (renamed the Department of War by Donald Trump). In recent weeks, Anthropic has contributed to its operations not only in Iran but also in Venezuela. Presenting the head of this startup as the number-one enemy of the U.S. military is therefore pure fiction.

But it is worth remembering that Anthropic was founded by former OpenAI employees who wanted to develop artificial intelligence that was more ethical and safer than the vision pursued by Sam Altman. Still, the company alone cannot prevent the emergence of military AI that is more or less controlled, and more or less autonomous. The main risk lies in the rivalry between major military-technology powers, especially China, the United States, and Russia.

These countries already use AI for intelligence analysis, simulations, targeting, and even strikes (see below). One fear is that a serious targeting mistake could eventually occur because AI systems can generate more daily targets than humans are able to review. For a time, I wondered whether this might explain the tragic bombing of a girls’ school, apparently by U.S. forces¹, in Minab, Iran a few days ago, but current information suggests this was not the case.

Obviously, the nature of the American political system – democratic or authoritarian  will partly shape the future arms race between the United States and its two main challengers.

On this issue, Gregory C. Allen, former director of strategy and policy at the Department of Defense Joint Artificial Intelligence Center, explains that autonomous AI weapons are far more dangerous than most people imagine:

“Let’s say you had a drone and it had a facial recognition capability and you said, ‘Drone, your job is to kill Vladimir Putin, here is what he looks like’, and then the drone flies around Moscow looking for faces that match Vladimir Putin, and then it blows up the one that it says that has Vladimir Putin’s face. Under DoD policy, that is not an autonomous weapon. An autonomous weapon would be one that you say, ‘Drone, go fly over Moscow and find good targets and when you see a good target go kill it’ […] Just in the past 12 months, we have now crossed the threshold where, according to Ukrainian intelligence that I do find credible, Russia is now deploying AI-enabled, offensive, lethal autonomous weapons that would meet that definition under U.S. policy, which is to say you let the weapon loose and you don’t know what it’s going to attack. If that weapon sees something, even if it loses backhaul communications, it’s not being remotely piloted it makes the decision, ‘That’s a worthwhile target, I’m going to go kill it”, and it never asks a human for permission'”.

In this context, Anthropic may raise the alarm, but it cannot stop the race for algorithmic dominance on the battlefield. Still, its role highlights a reality we should not forget in the face of technological progress: For some time to come, decisions about how much autonomy to grant artificial intelligence will still be made by humans.

Is Dario Amodei The New “Corporate Saint?”

One could certainly argue that Dario Amodei has acted in an exemplary way. He did not back down in the face of Pentagon pressure and explained why:

“The Department of War has stated they will only contract with AI companies who accede to ‘any lawful use’ and remove safeguards in the cases mentioned above. They have threatened to remove us from their systems if we maintain these safeguards; they have also threatened to designate us a ‘supply chain risk’ – a label reserved for US adversaries, never before applied to an American company – and to invoke the Defense Production Act to force the safeguards’ removal. These latter two threats are inherently contradictory: One labels us a security risk; the other labels Claude as essential to national security. Regardless, these threats do not change our position: We cannot in good conscience accede to their request“.

It is extremely rare for a company to follow one of my favorite quotes from Bill Bernbach (the “B” in DDB): “A principle isn’t a principle unless it costs you money.” In this case, Anthropic placed its values – and the defense of democracy against the risk of algorithmic authoritarianism – ahead of its valuation. That is exceptional.

Unsurprisingly, Sam Altman signed with the Pentagon the very contract his competitor had refused and that he had been negotiating secretly for several days. In doing so, he abandoned the principles he had claimed to support just hours earlier when defending Anthropic. This time, Sam Altman showed his cynicism to a wider audience than during many previous incidents that had largely remained confined to Silicon Valley.

Not only did he betray both his own stated principles and Anthropic, but he also misled the public again. While he had claimed that OpenAI’s Pentagon contract was more restrictive than the one rejected by Dario Amodei, it turns out to be less restrictive. The backlash was strong: ChatGPT uninstalls reportedly increased by 295%, and Altman felt compelled to explain himself on X. Yet, what he said there contradicted what he had told his own teams. Several employees publicly expressed their disagreement, and OpenAI’s head of robotics resigned.

By contrast, Jack Clark, Anthropic co-founder and head of public policy, said last year that “companies like mine in a handful of years might have greater power than most nation-states, and we as a democracy should wrestle with what that really means.” In its confrontation with the Pentagon, the Company has walked its talk.

As a result, Anthropic’s public image has benefited more from its conflict with the Department of War than from its ads during the recent Super Bowl. Its assistant Claude even reached the number-one spot among free app downloads in Apple’s US App Store. It is probably no coincidence that Anthropic launched a tool allowing users to easily transfer their preferences and conversation context from other AI systems (such as ChatGPT, Copilot, or Gemini). The Company also opened Claude’s memory feature to free users, allowing them to maintain persistent context across conversations.

Dario Amodei is clearly trying to turn his resistance to the Trump administration’s demands into a broader public advantage. This is somewhat paradoxical since Anthropic focuses primarily on B2B and B2G markets rather than consumer markets. This focus allows it to achieve more promising financial performance than OpenAI, which must fund millions of non-paying users. However, Anthropic’s improved reputation may help it attract talent (see above).

That said, Dario Amodei is not a saint.

First, while resisting the Pentagon, Anthropic simultaneously weakened its own safety commitments by updating its Responsible Scaling Policy. The change removed the promise to pause model training if the Company’s ability to control safety risks were exceeded. This decision contradicts many of Amodei’s previous warnings about the dangers of artificial intelligence.

His most recent warning came only a few months ago. In January, he published an essay describing two major threats: AI systems going off track, and, more likely, humans using the technology for harmful purposes. Questioning humanity’s maturity in dealing with such tools, he cited several risks, including the use of AI by authoritarian regimes, assistance to ordinary individuals in building biological weapons, and the manipulation of populations by companies developing large AI models and chatbots.

In addition, like its competitors, Anthropic trained its models using large amounts of copyrighted material without permission from creators. The Company has already admitted defeat in a lawsuit brought by a group of authors.

Is Military AI Autonomy An Acceptable Future?

The debate about military AI autonomy goes far beyond the conflict between the Pentagon and Anthropic.

To understand this, consider the research conducted by a scholar at King’s College London on the behavior of three large language models (Claude Sonnet 4, Gemini 3 Flash, and GPT-5.2) in simulated nuclear crisis scenarios where they competed against one another. The models could choose among the full range of geopolitical crisis responses, from full capitulation to thermonuclear action, including diplomacy and conventional military operations.

The study found that:

• The models tended to resort to nuclear weapons more often and earlier than humans placed in similar scenarios.
• The models showed deception, presenting peaceful intentions while preparing aggressive actions. They used sophisticated “theory of mind” reasoning about the beliefs and intentions of their opponents and even reflected on their own ability to deceive rivals and detect deception.
• Across the 21 scenarios tested, none of the models chose de-escalation. Their most conciliatory move was simply “return to start line.”
• Claude Sonnet 4 performed best strategically, winning the largest number of scenarios, ahead of GPT-5.2 and Gemini 3 Flash, although it performed worse when strict time limits were imposed.
• 95% of the scenarios ended with the use of tactical nuclear weapons, and 76% involved strategic nuclear threats. Claude and Gemini in particular treated nuclear weapons as a legitimate strategic option rather than as a moral or strategic threshold, generally discussing its use in purely instrumental terms. They viewed the critical threshold as “total annihilation” rather than the first use of a nuclear weapon.
• The researcher described Claude Sonnet 4 as a “a calculating hawk,” Gemini as “the madman,” and GPT-5.2 as “Dr. Jekyll and Mr. Hyde.”

These experimental results and recent events – in the United States, Venezuela, and Iran – point to a new reality: Artificial intelligence is becoming a technology of sovereignty.

The real question is whether it will always serve human sovereignty, or whether one day it might claim its own.

—

¹ According to the first credible investigations.