How Anthropic Took On The Pentagon Over The Future Of Military AI

Welcome to Memorandum Deep Dives. In this series, we go beyond the headlines to examine the decisions shaping our digital future. 🗞️

This weekend, we’re examining what happened when one of the world’s leading AI labs refused to give the Pentagon unrestricted access to its technology. What began as negotiations over a military contract quickly escalated into a confrontation between Anthropic and the U.S. government, ending with the company’s blacklisting and the Pentagon turning to a rival the very same day.

For years, artificial intelligence has spread rapidly through consumer products, enterprise tools, and digital platforms. But as the technology becomes increasingly central to national security, the relationship between the companies building AI and the governments seeking to deploy it is beginning to change. The Anthropic episode offers a rare look into that shifting dynamic, raising deeper questions about surveillance, autonomous weapons, and whether private firms can realistically set limits on how states use the most powerful technologies they create.

The standoff

For months, it had been a quiet negotiation. Anthropic had been in talks with the Pentagon over a contract reportedly worth up to $200M. The Department of Defense sought access to Claude, Anthropic’s AI model, for deployment across military applications, and Anthropic agreed. The frontier AI lab supported, as CEO Dario Amodei would later say, “all lawful uses of AI for national security.” But it wanted two things written into the contract that the Pentagon would not accept.

The first was a prohibition on using Claude for mass domestic surveillance of American citizens, and the second was a ban on integrating Claude into fully autonomous weapons systems, defined as those capable of selecting and striking targets without direct human oversight.

The conditions were not broad philosophical objections; they were specific contractual red lines that would establish defined guardrails on the use of the technology for military applications, which it considered technically dangerous and ethically unacceptable.

As Amodei wrote in his public statement, frontier AI systems are “not reliable enough to power fully autonomous weapons,” and deploying them that way “would endanger America’s warfighters and civilians.” On surveillance, he argued that AI’s capabilities are “getting ahead of the law,” making things possible that existing legal frameworks were never designed to constrain.

However, the Pentagon saw it differently, and from its perspective, an AI company dictating which government applications were acceptable set a precedent in which private firms held veto power over national security tools. The military wanted unrestricted access across all lawful use cases, no carve-outs, and no conditions.

By late February 2026, Anthropic rejected the Pentagon’s final offer, but what came next would make global headlines. Following the rejection, President Trump ordered all U.S. government agencies to stop using Anthropic’s products, and Pete Hegseth, the U.S. Secretary of War, labeled Anthropic a ‘supply chain risk,’ effectively blocklisting it from any government or military work.

The very same day Anthropic was blacklisted, OpenAI announced it had reached an agreement with the Pentagon to deploy its models within the military’s classified network. And while CEO Sam Altman said OpenAI shared the same ‘red lines’ as Anthropic, stating that they would restrict mass domestic surveillance, autonomous weapons, and high-stakes automated decisions such as social credit systems, the actual contract told a different story.

Critics quickly pointed out that OpenAI’s initial agreement did not explicitly prohibit the collection of Americans’ publicly available information. Cell phone location records, fitness app data, social media activity, commercial data broker files: all of these are technically publicly available. An AI system that aggregates them can build surveillance profiles functionally equivalent to wiretapping, without triggering existing legal protections.

The backlash to OpenAI’s agreement with the Pentagon was immediate and intense, with social media abuzz with campaigns to cancel ChatGPT subscriptions. The campaigns resulted in a spike in ChatGPT uninstallations and Claude rising to the number one spot in the U.S. App Store.

By early March, CEO Sam Altman was in damage control mode, publicly admitting the deal was “opportunistic and sloppy.” OpenAI further amended the contract, adding explicit language that the AI system “shall not be intentionally used for domestic surveillance of U.S. persons and nationals.” The revision also covered commercially purchased data, closing the most obvious loophole.

Whether the amended language is sufficient, and whether it will survive contact with the operational reality of a classified military network, are open questions. However, despite the lack of clarity, the incident showcased the growing tension between developers and deployers of AI tools.

The precedent, not the contract

To understand why the Anthropic episode matters beyond the news cycle, it helps to focus on what the Pentagon demonstrated. In the span of a week, the U.S. government showed it could block an AI company, label it a national security risk, cut off government revenue, and pivot to a competitor with little friction.

The message to the AI industry was clear: refusal carries consequences. And though Anthropic absorbed the loss of a contract and was rewarded by consumers for taking a principled stand, smaller firms watching this unfold may draw a different lesson: that compliance is safer than resistance.

The deeper issue, though, is not simply whether the military wants AI; that has been obvious for some time now. What matters now is how quickly and decisively the state can compel AI companies to align with its demands, which can easily slip into troubling water in the absence of clearly defined regulations.

The surveillance problem hiding in plain sight

Although much of the public debate has centered on autonomous weapons, the more immediate and consequential application of AI lies in intelligence and surveillance, where modern systems can process communications metadata, financial records, travel patterns, social media activity, and location data at a scale that removes the practical constraints that once limited state monitoring.

Legal frameworks such as the Fourth Amendment and the Foreign Intelligence Surveillance Act were written in an era when surveillance demanded sustained human effort and therefore carried natural limits on scope. Still, AI compresses the cost and effort required to monitor individuals and populations alike, raising questions about whether contractual promises against intentional domestic surveillance are meaningful in a world where pattern-recognition systems can synthesize vast data streams in real time.

The slow erosion of the human in the loop

Similarly, the concern over autonomous weapons reflects a longer-term anxiety, yet here too the change is likely to unfold gradually rather than through a single defining decision. AI is already woven into logistics, intelligence, and operational planning, and the real issue is not its presence but the steady migration of authority from human operators to machine systems as they prove faster and more effective, subtly shifting the human role from decision-maker to supervisor and eventually to nominal overseer. Each step may appear operationally rational on its own, but taken together, they risk transforming the substance of human control into something far thinner than the term implies.

A turning point, not a footnote

The episode ultimately revealed less about one company’s contract and more about the distribution of power between the state and the firms building foundational technologies. It shows that access to government markets can be conditioned swiftly and decisively, that classified deployments can expand beyond public scrutiny, and that the incentive structure increasingly favors alignment with state priorities over resistance to them. As the United States asserts this leverage, other governments acquire both precedent and justification to demand similar compliance from their domestic AI industries, accelerating the integration of advanced AI into national security infrastructures worldwide.

The unresolved question, which extends far beyond any single agreement, is whether societies can adapt their legal and oversight frameworks quickly enough to govern systems whose capabilities are advancing faster than the rules meant to contain them.

Transformative technologies do not wait for the rules designed to govern them. They move, they are adopted, and the reckoning comes later. The question the Anthropic episode leaves open is not whether AI will be integrated into military and state power, but whether the societies it serves will retain meaningful control over how it is used before that question becomes impossible to answer.