OpenAI Took the Pentagon Deal Anthropic Was Blacklisted For — Then Agreed to the Same Terms

On February 27, 2026, the Trump administration did something unprecedented in the history of American technology policy: it designated Anthropic — a US-based AI company — as a "Supply-Chain Risk to National Security." President Trump issued a directive ordering every federal agency to immediately cease all use of Anthropic's technology. Anthropic's $200 million Pentagon contract was cancelled.

Within hours, OpenAI announced it had signed an agreement to deploy its models in the Pentagon's classified network.

Within days, the full details emerged — and they contained one of the more remarkable ironies in recent tech history: OpenAI had agreed to the exact same safety conditions that Anthropic had been blacklisted for demanding.

This is the full story, and what it means for developers and the AI industry.

What Anthropic Was Punished For

To understand why the OpenAI deal matters, you need to understand what triggered Anthropic's blacklisting.

Anthropic had been in contract negotiations with the Pentagon for months. The sticking point was specific: Anthropic insisted on explicit contractual prohibitions on using Claude for:

1. Mass surveillance of American citizens
2. Fully autonomous weapons systems

These weren't abstract ethical principles — they were specific, documented contract terms that Anthropic refused to remove. Defense Secretary Pete Hegseth characterised this as Anthropic trying to "seize veto power over the operational decisions of the United States military." The administration responded by making Anthropic the first American company ever to be designated a national security supply chain risk — a designation typically reserved for adversarial foreign entities.

The message was clear: comply or be cut off from federal business entirely.

OpenAI's Move

OpenAI moved fast. The deal was announced Friday evening, February 28 — less than 24 hours after the Anthropic blacklisting. Sam Altman announced on social media that OpenAI had reached agreement with what he called "the Department of War" (the Trump administration has signalled a potential renaming of the Department of Defense) to deploy models in their classified network.

The announcement was notable for what Altman said next — he described the safety conditions OpenAI had secured:

> "Two of our most important safety principles are prohibitions on domestic mass surveillance and human responsibility for the use of force, including for autonomous weapon systems. The DoW agrees with these principles, reflects them in law and policy, and we put them into our agreement."

Read that again slowly. No domestic mass surveillance. No autonomous weapons. Human responsibility for use of force.

These are precisely the conditions Anthropic was blacklisted for demanding 24 hours earlier.

The Irony, Stated Plainly

The Trump administration designated Anthropic a national security threat for insisting on safety conditions. Then it accepted identical safety conditions from OpenAI.

The difference was not the substance of the safety requirements. The difference appears to have been the manner of negotiation — and possibly the commercial and political relationships involved.

Altman himself acknowledged the situation was awkward: "Definitely rushed, and the optics don't look good." He stated OpenAI moved quickly partly to "de-escalate" what was becoming an increasingly heated standoff between the US military and the AI industry.

What OpenAI Actually Agreed To

The agreement includes four specific safety conditions:

1. No domestic mass surveillance

No use for unconstrained monitoring of US persons' private information without legal authorisation.

2. No autonomous weapons control

No use to direct autonomous weapons systems. Humans retain responsibility for decisions involving the use of force.

3. No automated high-stakes decisions

No use for automated decisions analogous to "social credit" systems that affect individuals without human review.

4. Law enforcement restrictions

No use for domestic law enforcement except as permitted by the Posse Comitatus Act — the federal law that restricts military involvement in civilian law enforcement.

The Technical Implementation

Beyond the contractual terms, OpenAI secured specific technical protections that distinguish this from a policy-only agreement:

Cleared engineers on-site: Forward-deployed OpenAI engineers with security clearances are stationed at the Pentagon to monitor implementations and ensure model safety. This is significant — it means OpenAI has visibility into how its models are actually being used in the classified environment.

Safety stack preserved: OpenAI retains full discretion over its safety stack. The Pentagon cannot require OpenAI to remove or modify safety guardrails as a condition of deployment.

Alignment researchers in the loop: OpenAI's safety and alignment researchers are kept informed of classified deployments. The safety team is not ringfenced from the military work.

Cloud-only, no edge devices: The deployment is cloud-based, not deployed to edge devices or weapons systems. This keeps the physical implementation within the more controllable cloud environment.

Altman noted that Pentagon officials were "genuinely surprised we were willing to consider classified work" — suggesting the DoD expected OpenAI to decline, as several other companies have.

The Backlash: #CancelChatGPT

The deal triggered immediate and significant public reaction, particularly within the tech and AI communities.

Many OpenAI employees signed an open letter expressing solidarity with Anthropic's stance and opposing the terms of engagement with the Pentagon. The irony of OpenAI's position — having agreed to the same conditions Anthropic was punished for — did not provide complete cover from criticism. Critics argued that regardless of the safety terms, the deal normalises AI companies doing classified military work at all.

A "#CancelChatGPT" boycott movement emerged on social media, urging users to switch from ChatGPT to Anthropic's Claude in solidarity.

The market response was concrete: Claude surged to become the most downloaded free app in Apple's App Store in the days following the deal — a direct result of users following through on the boycott. It was, in a strange way, the best advertising Anthropic could have received: their blacklisting for principled refusal drove a measurable user acquisition wave.

Emil Michael, Under Secretary of Defense for Technology, defended the agreement: "When it comes to matters of life and death for our warfighters, having a reliable and steady partner that engages in good faith makes all the difference as we enter into the AI Age."

What This Means for AI Developers

This episode clarifies several things that were previously uncertain about the frontier AI industry:

1. Safety terms are negotiable — and can be contractual

The conventional wisdom was that government/enterprise AI deployments required AI companies to strip safety guardrails on request. The Pentagon-OpenAI deal demonstrates that safety conditions can be contractual, technical (not just policy), and accepted by government clients. This has implications for how enterprise AI contracts are structured across the industry.

2. The difference between Anthropic and OpenAI wasn't the terms — it was the politics

Both companies wanted essentially the same safety protections. One was blacklisted; the other got a contract. The lesson for AI companies navigating government relationships: the substance of your position matters less than the political packaging and the relationships involved.

3. Cleared engineers as a safety mechanism is a new model

Embedding cleared AI company engineers inside classified deployments is a novel approach to safety oversight. For developers thinking about how to maintain safety visibility in sensitive deployments, this forward-deployment model is worth watching.

4. The boycott effect is real

Claude reaching #1 in the App Store from a political boycott is a data point developers building AI-adjacent products should note: AI model choice is increasingly a values-based consumer decision, not just a technical one. Users are paying attention to the ethics and political positioning of the AI companies whose models power the products they use.

5. OpenAI is now inside the classified national security apparatus

This is a structural change in the AI landscape. OpenAI models are now running on classified Pentagon infrastructure. The implications for the long-term development trajectory of these models, and for competitors who choose not to participate in military work, will take years to fully understand.

The Bigger Picture: AI in Military Contexts

The Anthropic-Pentagon-OpenAI sequence is not an isolated incident. It reflects a broader pattern: the US government is actively integrating frontier AI into national security operations, and the terms on which AI companies participate are being established right now.

The stakes are high on multiple sides. For AI safety advocates, the concern is that competitive pressure will push companies to accept military contracts with progressively weaker safety conditions. The OpenAI deal provides partial reassurance — but only because the conditions were accepted this time.

For developers building on top of OpenAI APIs, the deal is relevant context: the company whose models you're building with is now operating in classified national security environments. This affects how you should think about API reliability, policy changes, and the long-term strategic direction of the platform.

For the broader tech industry, the Anthropic blacklisting — the first time an American tech company was designated a national security supply chain risk — sets a precedent with potentially wide reach. The designation was used as apparent leverage in a contract negotiation. If this becomes an accepted tool of government procurement, it changes the power dynamics of every AI company doing government work.

Related: Anthropic Pentagon Blacklisted: Supply Chain Risk Designation Explained