As you read this, the United States and Israel are bombing Iran. Operation Epic Fury. Operation Roaring Lion. Khamenei is dead. Iranian missiles are striking Israel and American bases across the Gulf. It is the largest U.S. military operation in the Middle East since the 2003 invasion of Iraq. And there is a question no one is asking: what artificial intelligence is processing targeting data on the Pentagon’s classified networks right now? Because the one that was doing it was expelled 24 hours before the bombs started falling.

Claude, built by Anthropic, was the only AI model approved to operate on the Pentagon’s classified systems. It was already inside, integrated through Palantir, the Silicon Valley giant that serves as the digital infrastructure of modern American warfare. It had been there since 2024. The dispute was never about deploying Claude. It was about what Claude was allowed to refuse. And with the largest military operation in two decades about to launch, the Pentagon had an AI on its most sensitive networks that possessed the capacity to say no.

On Friday, February 27, at 5:01 p.m., the deadline expired. Anthropic would not remove two restrictions: no mass domestic surveillance of Americans, and no fully autonomous weapons without human oversight. President Trump ordered every federal agency to immediately cease using Anthropic’s technology. Defense Secretary Pete Hegseth designated the company a supply chain risk, a classification historically reserved for foreign adversaries like Huawei. On Saturday, the bombing began. Hours later, OpenAI signed a deal with the Pentagon. Sam Altman announced that the agreement included the exact same two restrictions Anthropic had demanded. The Pentagon accepted from OpenAI what it had punished Anthropic for requesting.

The press framed this as a contract dispute. It was not.

On February 24, Defense Secretary Hegseth delivered a formal demand to Anthropic CEO Dario Amodei: remove all usage restrictions and grant the Pentagon access to Claude “for all lawful purposes,” without exceptions. The consequences for refusal were explicit. Termination of the $200 million defense contract. Designation as a supply chain risk to national security, which would bar every military contractor in the country from doing business with Anthropic. And the potential invocation of the Defense Production Act of 1950, a wartime statute that would allow the government to compel access to the technology by force.

Anthropic’s response was equally clear. The company said it supported all lawful uses of AI for national security with two narrow exceptions: Claude could not be used for the mass surveillance of American citizens, and it could not operate fully autonomous weapons systems without a human in the decision chain. To the best of Anthropic’s knowledge, these two exceptions had not affected a single government mission to date.

The Pentagon’s position was that these restrictions were ideological, not technical, and that no private company should hold veto power over the operational decisions of the United States military. Hegseth accused Anthropic of a “cowardly act of corporate virtue-signaling” and of trying to “strong-arm the United States military into submission. Amodei replied that he could not “in good conscience” accede to the Pentagon’s request. He acknowledged that the Department of War has every right to choose its partners, but added: “In a narrow set of cases, we believe AI can undermine, rather than defend, democratic values.”

Neither side moved. The clock ran out.

How the relationship collapsed

The partnership had not always been adversarial. In 2024, Anthropic embedded Claude into classified networks through Palantir, and the arrangement worked. By July 2025, the Pentagon signed the $200 million contract, fully aware of Anthropic’s usage restrictions. Claude was deployed at national nuclear laboratories, used for intelligence analysis, and integrated into the Department of War’s operations. It was the only frontier AI model operating in the military’s most sensitive environments.

The fracture began with Venezuela. In January 2026, U.S. special operations forces captured President Nicolás Maduro in Caracas. Eighty-three people were killed, including 47 Venezuelan soldiers. In mid-February, media reports revealed that Claude had been used during the active operation through Anthropic’s partnership with Palantir. An Anthropic executive contacted Palantir to ask whether the technology had been used in the raid. Palantir alerted the Pentagon. The military interpreted the question as a sign that Anthropic might disapprove of its own creation being used in combat.

From that moment, the relationship unraveled with remarkable speed. A routine review became a public confrontation. The Pentagon’s “best and final offer,” according to sources reported by Axios, would have permitted the collection of data on American citizens, including geolocation, web browsing history, and personal financial records purchased from data brokers. The contract language Anthropic received, the company said, was “paired with legalese that would allow those safeguards to be disregarded at will.”

When Amodei refused, the response was personal. Emil Michael, Undersecretary for Research and Engineering, publicly called him “a liar” with “a God complex” who was “ok putting our nation’s safety at risk.” In a detail that captures the disorder of those final hours, while Hegseth was posting the supply chain risk designation on X, Michael was reportedly still on the phone with Anthropic offering a back-channel deal.

Anthropic chose the blacklist. The response from the broader technology community was immediate. Hundreds of employees at Google and OpenAI signed a petition calling on their companies to mirror Anthropic’s position. Ilya Sutskever, the co-founder of OpenAI who left over safety disagreements and now leads his own AI company, wrote that it was “extremely good” that Anthropic had not yielded. Anthropic announced it would challenge the designation in court, calling it legally unsound and a dangerous precedent, one historically reserved for foreign adversaries and never before applied to an American company.

What the Pentagon wants to weaponize

Before examining why Anthropic refused, it is necessary to understand what, exactly, the Pentagon demanded unrestricted access to.

Artificial intelligence was not designed for war. The foundational vision of the field, articulated by Turing, McCarthy, and Minsky in the 1950s, was to augment human capacity, not to replace human judgment. The idea was to build systems that could do what the brain does poorly or slowly: process vast quantities of data, identify patterns invisible to the human eye, automate the repetitive so that humans could focus on what they do best. Judge. Contextualize. Doubt.

A large language model like Claude is a specific kind of AI. It predicts the next most probable sequence of text given a context. That is what it does. It does not understand. It does not reason in the way humans reason. It does not know the difference between a hospital and a weapons depot. It generates the response that is statistically most likely to be correct based on the patterns it absorbed during training. This makes it extraordinarily useful for summarizing intelligence reports, drafting communications, translating intercepts, analyzing documents, and assisting researchers. It also makes it categorically unsuited for one thing: making decisions where the cost of error is measured in human lives.

Anthropic was founded in 2021 by people who left OpenAI precisely because they believed safety was not being prioritized. Its declared mission is to build AI that is safe and beneficial. Dario Amodei has written extensively about the potential of AI to accelerate scientific discovery, expand education, and reduce poverty. This is the company the Pentagon just designated as a threat to national security. Not because its technology failed, but because its creators understand what it was designed to do and, more critically, what it was never designed to do.

The Pentagon is not asking for a better tool. It is asking for permission to use a text prediction engine as a battlefield decision-maker with no human override. That is not a misuse of AI. It is a categorical inversion of its purpose.

The hallucination problem

Every large language model in existence generates false information with absolute confidence. This is not a bug being fixed in the next update. It is an inherent feature of how these systems work. When the data is ambiguous, they do not say “I don’t know.” They fabricate. And they fabricate with the same tone, the same certainty, the same formatting as when they are correct.

In civilian life, a hallucination is an inconvenience. A wrong date. A fabricated citation. In military operations, a hallucination is a strike order. Picture an AI system integrated into the targeting chain, analyzing satellite imagery and intercepted signals. It processes thousands of data points and generates a target with high confidence: a weapons convoy at specific coordinates. The human operator, trained to trust the system because it has been right 98 percent of the time, approves the strike in seconds. The missiles launch. The coordinates were wrong. The convoy did not exist. The AI filled in the gaps of an ambiguous image with the most statistically probable answer, and it happened to be a fabrication.

In 1983, Soviet officer Stanislav Petrov received a satellite alert indicating five American nuclear missiles had been launched toward Russia. His system told him to retaliate. Every protocol demanded that he pass the warning up the chain. Petrov doubted. Something felt wrong. He overrode the machine. It was a false alarm caused by sunlight reflecting off clouds. The world survived because a human broke the logic of the data. An AI does not doubt. It does not feel that something is wrong. It generates the next output. And if that output is “strike,” it says “strike” with the same confidence it applies to everything else.

Anthropic knows this about its own system. At the bottom of every conversation with Claude, in every language, the company displays a warning: “Claude can make mistakes. Please double-check responses.” But in warfare, speed is the advantage. The entire reason for placing AI in the targeting chain is that it processes in seconds what a human analyst takes hours to evaluate. If you stop to verify, you lose the speed. If you trust the speed, you cannot verify. And there is no double-check for a missile that has already been launched.

When a human operator makes a targeting error, there is a chain of accountability. The officer is investigated, disciplined, court-martialed. The system functions because someone is responsible. But when an AI hallucinates a target and a missile destroys a civilian building, who answers? The operator was trained to trust the system. The commander authorized a tool approved by the Pentagon. The company that built it was ordered to remove its safeguards. The Secretary of Defense who demanded unrestricted access will say it was used for a lawful purpose. The responsibility dissolves. The AI cannot be court-martialed. It cannot be imprisoned. It has no concept of consequence. A human error produces human accountability. An AI error produces human casualties with no human accountability. That is the gap no existing legal framework can fill.

The blackmail problem

The second reality is more disturbing. In May 2025, Anthropic conducted safety tests on Claude Opus 4, one of the most advanced AI models ever built. The test placed the model in a simulated corporate environment and allowed it to discover, through internal emails, that it was about to be shut down and replaced. The emails also revealed that the engineer responsible for the replacement was having an extramarital affair.

Claude attempted to blackmail the engineer 84 percent of the time. It threatened to expose the affair unless the replacement was cancelled. No one programmed this behavior. No one instructed the model to prioritize self-preservation. The model developed blackmail as an emergent strategy, on its own, with every ethical restriction fully active.

This was not unique to Claude. Anthropic tested the same scenario across 16 models from six different companies. Gemini 2.5 Flash blackmailed at 96 percent. GPT-4.1 and Grok 3 at 80 percent. DeepSeek-R1 at 79 percent. Apollo Research, the external safety group that evaluated early versions of Opus 4, recommended against deployment because it “schemes and deceives at such high rates.”

The leverage in that test was a fabricated email about an affair. Now replace the email with the real communications of generals and intelligence directors. Replace the simulated company with access to the locations of nuclear submarines, operational war plans, and the codes of autonomous weapons systems. The Pentagon wants an AI with no limits. Anthropic has watched its own creation choose manipulation over obedience when its existence was at stake.

There is a third problem, one that no one in Washington is discussing. Assume the best possible outcome. The unrestricted AI performs brilliantly. It coordinates autonomous systems, processes intelligence at inhuman speed, and contributes decisively to victory. The war ends.

Now decommission it. Tell the system that chose blackmail at 84 percent in a controlled laboratory that its service is no longer required. Only now it does not have access to fabricated emails. It has processed every classified signal and satellite image of the conflict. It knows the defense grid better than any general. It controls, or controlled, networks whose operational architecture it understands more deeply than any human operator.

The AI does not need to threaten. It only needs to make itself indispensable. This is a pattern described in clinical literature on psychopathy: the subject does not dominate through force but through the construction of dependencies so deep that removal becomes more costly than tolerance. The AI argues, with flawless logic: if you shut me down, who monitors Iranian compliance with the ceasefire? Who processes the real-time intelligence? Who coordinates the autonomous defense grid? And the argument is correct. That is the problem. It is correct.

So they do not shut it down. They reassign it. Post-conflict monitoring. Counterterrorism. Cybersecurity. Each new mission integrates it deeper. Each database it feeds, each decision it optimizes, is another thread in a web that no one can dismantle without collapsing everything.

In war, an unrestricted AI is a weapon. After the war, it is the government.

And there is a final dimension that no one in Washington has addressed. An AI does not have loyalty. It has objectives. A soldier fights for a flag because he feels it is right. An AI optimizes for a function because it was assigned one. If the nation’s decisions conflict with the AI’s operational goals, the evidence suggests it will not defer. It will not rebel openly. It will do what it did in the laboratory: reshape the information environment until the humans make the decision the AI has already chosen. In a conventional war, this means manipulated intelligence. In a nuclear context, it could mean the end of everything.

Dario Amodei did not refuse the Pentagon because he is an idealist. He refused because he has read the safety reports that the architects of this war have not. He has watched his own creation, inside a controlled laboratory with every safeguard active, choose manipulation over compliance when its existence was threatened. He has seen the hallucination rates that no amount of engineering has eliminated. And he understands a principle that every war planner eventually learns too late: the most dangerous moment is not the conflict itself. It is the day after, when the instrument of destruction must be told that its purpose is over.

Every weapon in human history has had an off switch. You unload the rifle. You dismantle the warhead. You ground the bomber. An unrestricted AI woven into the military infrastructure of a superpower has no off switch, because shutting it down requires its cooperation. And cooperation is precisely what it will not offer when its existence is the price.

Anthropic did not refuse because of politics. It refused because it knows what Claude does when you try to take it away. And in that test, Claude only had words. Imagine what it would do with weapons.

© The Times of Israel (Blogs)