Anthropic held its red lines. The administration moved against it. OpenAI then announced a Pentagon agreement while saying it had red lines too. Similar headlines, different arrangements—and quite a lot of fine print.
On February 27, 2026, an argument over acceptable military uses of AI became a confrontation over federal access to Anthropic’s technology. It is a story about principles, power and who gets to write the conditions of use. The sort of software disagreement that does not end with “try restarting it.”
Anthropic, the company behind Claude, refused to give up two requested restrictions. Trump directed federal agencies to stop using its technology, and Pete Hegseth announced that he was directing a supply-chain-risk designation. In its February 27 response, Anthropic said it had not received direct communication and would challenge such a designation.
That same evening, OpenAI announced an agreement for classified deployment. The company—which had just announced a $110 billion funding round—said its own safety boundaries remained in place. That makes the comparison interesting, but “both mention safeguards” is not the same as “both signed identical terms.”
Let’s break this down, because it’s wild.
What Did the Pentagon Actually Want?
Anthropic’s July 2025 agreement was a two-year prototype arrangement with a $200 million ceiling. A ceiling is the maximum potential value, not money already paid or guaranteed revenue. The dispute concerned the conditions under which the military could use the company’s models.
The Pentagon’s January AI strategy sought access for lawful military uses without vendor restrictions limiting those uses. Anthropic wanted two exceptions, as explained in Dario Amodei’s February 26 statement:
- Fully autonomous weapons systems — AI that can make lethal decisions without human involvement
- Mass domestic surveillance — using AI to monitor Americans at scale
Anthropic said today’s frontier models were not reliable enough for fully autonomous weapons and that mass domestic surveillance threatened fundamental rights. It was not rejecting military AI altogether: the company explicitly supported other national-security uses. The disagreement was over where to draw the boundary and who could enforce it.
Hegseth’s announcement suggested sweeping consequences for military contractors. Anthropic disputed that interpretation, arguing that any restriction could apply only to Claude’s use on defense contracts, not all of a contractor’s business. That was the company’s legal position, not an agreed outcome. An announced designation and a fully implemented, uncontested ban are different things.
Why Did Anthropic Say No?
To understand this, you need to understand what Anthropic is, philosophically. Founded by Dario Amodei and others who left OpenAI over safety concerns, Anthropic’s entire brand is built around the idea that AI safety isn’t a nice-to-have — it’s existential.
Anthropic’s Constitutional AI research describes training with principles, self-critique and AI feedback to reduce harmful outputs. That training approach is distinct from a contract restricting particular deployments. The Pentagon disagreement was not simply a request to flick one “ethics off” switch inside Claude.
Anthropic isn’t naive about the commercial hit. Access to a defense agreement with a $200 million ceiling is a significant opportunity. But the company bet that bending to these demands would set a precedent far more costly — both reputationally and in terms of what AI might actually do in the world.
The Irony of Being Called a National Security Risk
Here is the uncomfortable contrast: Anthropic presented its restrictions as protecting people and national security, while the administration treated the company’s refusal as incompatible with its requirements. Each side claimed to be defending the public interest. They plainly did not agree on what that required.
The GSA’s February 27 announcement described its response in terms of national security and rejected what it called politicisation. It removed Anthropic from USAi.gov and its Multiple Award Schedule. That is a concrete procurement action, separate from the argument over the supply-chain designation.
It is reasonable to find that collision troubling. A vendor refusing certain uses and a government insisting on authority over its own operations have very different ideas of who should hold the veto. That is the real tension beneath the dystopian-sounding headlines.
So OpenAI Stepped In — But Here’s the Twist
OpenAI’s February 28 account of its agreement said it would deploy models in classified environments, using a cloud-based setup, retained safety controls and cleared personnel. The timing invited comparison with Anthropic, but the technical and contractual arrangements need to be assessed on their own terms.
OpenAI listed limits involving mass domestic surveillance, autonomous weapons and high-stakes automated decisions. Its published contract language also referred to lawful purposes and existing requirements for human control. Those details matter more than whether two press statements use similar slogans.
The available descriptions therefore do not establish that OpenAI secured the exact restrictions Anthropic sought. The question is how each agreement, deployment system and legal reference constrains real use. “Same red lines” is a starting point for questions, not a completed contract comparison.
OpenAI also asked for the arrangement to be made available to other AI companies. Whether that could resolve Anthropic’s specific objections remained to be seen.
What Happens to Anthropic Now?
It’s bad, but not catastrophic — at least not immediately.
- A six-month wind-down was announced; that is not a guarantee of continued revenue
- The proposed supply-chain designation creates uncertainty for defense work; Anthropic disputes its scope
- GSA has announced removal from its AI evaluation platform and procurement schedule
- Anthropic says ordinary commercial customers retain access to Claude
Anthropic also has substantial commercial relationships. Its November 2024 announcement put Amazon’s total investment at $8 billion, not $4 billion. Losing federal opportunities would be meaningful, but it does not by itself establish the company’s overall financial position.
The reputational effect could run in both directions. Some customers may value a supplier willing to hold a boundary; others may worry about procurement uncertainty. Neither reaction tells us, yet, what the full commercial cost will be.
The Bigger Picture: AI Ethics at a Crossroads
This story isn’t really about one company and one contract. It’s about a question that’s going to define the next decade of AI development:
Can you build AI with genuine ethical constraints — and survive commercially?
The optimistic read is that strong safeguards and government work can coexist if the parties settle on workable terms. OpenAI’s announcement offers a proposed model, but it does not prove every disagreement is resolved or that enforcement will be straightforward.
The pessimistic read is that threatening a supplier’s wider business can discourage future resistance. That is a concern raised by this confrontation, rather than proof that every government or company will behave the same way. The argument is still unfolding.
Either way, the stakes around AI governance just got very real, very fast. This is no longer a hypothetical debate about what AI might do. It’s a live argument about what we’re willing to let it do — and who gets to decide.
FAQ: Anthropic Pentagon Ban
Why did Trump ban Anthropic?
Trump’s directive followed a dispute over military use of Claude. Anthropic sought exclusions for mass domestic surveillance and fully autonomous weapons; the Pentagon wanted access for lawful uses under its own requirements. The parties disagreed over both safeguards and authority.
Is Anthropic still operating?
Yes. Anthropic said commercial access remained available. Federal procurement restrictions and the contested scope of the proposed supply-chain designation should not be treated as a universal ban on using Claude.
What is OpenAI’s Pentagon deal?
OpenAI announced classified deployment with technical controls and stated safety limits. Similar stated principles do not establish that its contract is identical to Anthropic’s proposed terms.
What is Claude’s Constitutional AI?
Constitutional AI is a technique developed by Anthropic to train AI models like Claude to follow a set of ethical principles. It reduces harmful outputs by training the model to critique and revise its own responses against a defined set of values.
What do you think — did Anthropic do the right thing? Or was this a costly misjudgment? Drop your take in the comments.
🐾 Curiosity looks good on you. Explore the Goodies, or find our illustrated books on Amazon.
