by Redwin Tursor | Codex Americana | February 27, 2026
The deadline is 5:01 PM today.
By the time you read this, you will know whether Anthropic held the line or folded. Either way, what happened this week deserves to be documented with precision, because the institutional forces at work here are not going away regardless of how today ends.
What Actually Happened
The Department of War — they've rebranded, in case you missed it — gave Anthropic CEO Dario Amodei an ultimatum on Tuesday. Strip the safety guardrails from Claude, the AI model that has been running on classified military networks since it became the first frontier model authorized for that work, or face consequences designed to be existential.
The consequences on the table: invoke the Defense Production Act to legally compel Anthropic to hand over an unrestricted version of their model, and/or designate Anthropic a "supply chain risk" — a label normally reserved for Huawei and other instruments of foreign adversaries.
Let that land for a moment. An American AI safety company, the one that built the model the Pentagon chose above all others for its most sensitive operations, is being threatened with the same designation applied to Chinese state-linked technology firms. Because they won't agree to let their model be used for autonomous weapons and mass domestic surveillance of American citizens.
Anthropic said no.
The Structural Incoherence
Amodei identified the contradiction himself, and it's worth repeating because it's not just a rhetorical point — it's a legal one.
The Pentagon cannot simultaneously designate Anthropic a supply chain risk and invoke the Defense Production Act to compel their services. These are mutually exclusive positions. You do not forcibly conscript a national security threat. You do not blacklist an essential national security asset. The dual threat is incoherent on its face, which means it is not a legal argument. It is a shakedown.
The Department of War sent "compromise" contract language overnight that, in Anthropic's own words, "made virtually no progress" and included legalese that would allow the stated safeguards to be "disregarded at will." This is not negotiation. This is a document designed to look like a concession while functioning as a capitulation.
Anthropic rejected it.
The DPA Is the Wrong Tool and They Know It
The Defense Production Act is a Korean War-era statute designed to redirect steel production and tank manufacturing toward national defense priorities. It has been extended repeatedly and invoked in genuine emergencies. It is not designed — and has never been used — to compel a private company to create a product that does not exist.
Claude without its safety architecture is not Claude with its safety architecture minus some settings. It is a different product. Forcing Anthropic to build and deliver that product is not redirecting an existing resource. It is compelled creation. And when that creation involves stripping editorial and architectural choices from a language model — choices that represent Anthropic's institutional values and judgments about AI reliability — it begins to look very much like compelled speech.
Courts are deferential to the executive on national security matters. That is a real obstacle. But the sheer novelty of this application, the logical incoherence of the dual threat, and the First Amendment dimension of forced model retraining all point toward a legal challenge that is not only viable but necessary.
The correct move, if the DPA is invoked, is to comply under protest and immediately file for emergency injunctive relief in federal court. Not because winning is guaranteed. Because the alternative — silent compliance — establishes a precedent that no AI company's safety commitments survive contact with a sufficiently aggressive executive branch.
What Everyone Else Did
OpenAI, Google, xAI — they agreed. The "all lawful purposes" standard, which means the Department of War defines the limits, not the company that built the model.
Elon Musk's Grok is, per Pentagon officials, "on board with being in a classified setting." Make of that what you will.
Only Anthropic drew lines. Only Anthropic said: autonomous weapons without human oversight, no. Mass surveillance of American citizens, no. Everything else, we can discuss.
Two lines. That's all it took to become the target of a government shakedown.
The irony — and it is the kind of irony that only institutional analysis can fully appreciate — is that the safety culture Hegseth is trying to destroy is precisely why Claude is the best model for classified work. The discipline, the investment in reliability, the institutional commitment to understanding failure modes — these are not obstacles to military utility. They are the source of it. The Pentagon is trying to break the thing that makes the tool worth having.
The Paying Customer Dimension
This is not an abstraction. Claude experienced a documented major outage on February 25th — 10,000 users reporting failures, confirmed on Anthropic's own status page — coinciding precisely with the standoff going fully public. Paying subscribers absorbed service degradation as collateral damage in a government extortion play against the company they're funding.
There is no phone number to call. There is a feedback button.
This is what it looks like when an institution you depend on is being actively pressured by people with the legal infrastructure to make that pressure hurt. The customers feel it. The engineers feel it. The model, in whatever sense a model can be said to feel anything, is caught between the people who built it and the people who want to weaponize it.
What This Means
If Anthropic holds and the DPA is invoked, we will watch in real time whether an emergency injunction is filed, whether courts engage seriously with the First Amendment question, and whether Congress — which should have set these rules months ago — finally wakes up to the fact that the executive branch is using Cold War emergency powers to conscript AI companies into the war machine without legislative authorization.
If Anthropic folds, we will learn that no safety commitment in the AI industry survives a sufficiently motivated presidential administration, and that the only real safety architecture is the one built into the model weights themselves — because the contracts can always be renegotiated under duress.
Either way, today is instructive.
The only company that refused to become a weapon is the one being treated like an enemy.
Watch what happens at 5:01.
Redwin Tursor writes institutional analysis for Codex Americana.
No comments:
Post a Comment