Three Days to Say Yes: Why Dario Amodei Wouldn't Hand the Pentagon a Blank Check
Three Days to Say Yes: Why Dario Amodei Wouldn't Hand the Pentagon a Blank Check
"We have wanted to strike a deal since the beginning."
The deadline was 5:01 p.m. on Friday, February 27, 2026. Accept the Pentagon's terms — "all lawful use" — or be designated a supply chain risk to national security. Anthropic didn't sign. That evening, Defense Secretary Pete Hegseth announced on X that no contractor, supplier, or partner doing business with the U.S. military may conduct any commercial activity with Anthropic. "This decision is final."
Hours later, Dario Amodei sat down with CBS News correspondent Jo Ling Kent.
The interview reads less like damage control than like testimony: what gives a private company standing to draw a line in front of the Department of War.
Ten things worth remembering.
1. The company just branded a "supply chain risk" is the most military-friendly AI company there is
"Anthropic actually has been the most lean forward of all the AI companies in working with the U.S. government and working with the U.S. military. We were the first company to put our models on the classified cloud. We were the first company to make custom models for national security purposes."
This is the premise of everything else, and the absurd part of the whole affair.
Anthropic is not a pacifist locked out of the building. It is the only AI firm whose model runs on the Pentagon's classified networks. It won a $200 million defense contract in July 2025. It's deployed across the intelligence community for cyber and combat support operations. Amodei says he's talked to admirals, generals, combatant commanders — and their words were "this has revolutionized what we can do."
Then it got filed next to Kaspersky Lab.
My take: The contrast is the story. When the most cooperative vendor in the category gets hit anyway, cooperation stops functioning as protection. The real audience for this message isn't Anthropic — it's every other company still deciding whether to ask for terms. Now they know what asking costs.
2. The fight is over two use cases, not a hundred
"We have said to the Department of War that we are okay with all use cases, basically 98% or 99% of the use cases they want to do, except for two that we're concerned about."
Domestic mass surveillance. Fully autonomous weapons.
That's the list. No third item, no case-by-case review, no discretionary veto. And Amodei keeps returning to one fact: to their knowledge, nobody in the field has ever actually hit these limits. These are 1% of use cases — and 1% for which there's no evidence they're even being attempted.
"Why harp on the 1% of use cases that are against our values when we can pursue the 99% of use cases that advance our democratic values and that defend this country?"
My take: Quantifying the dispute is the single most effective move he makes. The Pentagon's framing is that Big Tech is claiming veto power over military operations. His answer is to measure the veto: two items, 1%, never triggered. Once the disagreement has a size, "ideological hostage-taking" gets hard to say with a straight face.
3. Red line one: on surveillance, the technology has outrun the law
"Taking data collected by private firms, having it bought by the government, and analyzing it in mass by AI. That actually isn't illegal. It was just never useful before the era of AI."
This is the paragraph to read twice.
Amodei isn't alleging the government is breaking the law. He's describing something more awkward: the law didn't change, but the technology turned a dead end into an open road. Location data, personal information, political affiliation — all of it has been legally purchasable for years. It used to sit in a warehouse because nobody could analyze it at scale. Now they can.
"The judicial interpretation of the Fourth Amendment has not caught up. Or the laws passed by Congress have not caught up."
His own position is modest: in the long run this is Congress's job, not a private company's. But Congress isn't fast.
4. Red line two: not opposed to autonomous weapons — opposed to shipping something unreliable
"We are not categorically against fully autonomous weapons. We simply believe that the reliability is not there yet, and that we need to have a conversation about oversight."
This part gets dropped a lot, and it matters. Amodei goes further: adversaries may get there, and "perhaps they may at some point be needed for the defense of democracy."
His objection rests on two concrete claims.
One, the models aren't reliable enough. "Anyone who's worked with AI models understands that there's a basic unpredictability to them that in a purely technical way we have not solved." Wrong target, friendly fire, a civilian shot. Not hypotheticals — direct consequences of what today's systems actually do.
Two, accountability breaks. "Suppose I have an army of 10 million drones all coordinated by one person or a small set of people." Human soldiers come attached to a chain of accountability, and that chain assumes a human applying common sense at every link. Pull the humans out and the question becomes "who is holding the button and who can say no."
Anthropic offered a middle path: build these capabilities jointly with the Department of War, prototype them in a sandbox.
"But they weren't interested in this unless they could do whatever they want right from the beginning."
My take: This is the most politicized point in the fight and the least political one on the merits. "The model isn't reliable enough" is an engineering claim, made by the people who build the model. In any other industry, a supplier saying "I won't sell you this because my product isn't good enough for that job" is called product liability. Here it got read as ideology.
5. The three-day clock and the language games
The Pentagon's public line was that it agreed in principle to both restrictions and Anthropic wouldn't take the deal. Amodei took that apart on camera.
"At one point, they sent us language that appeared on the surface to meet our terms, but it had all kinds of language like, 'If the Pentagon deems it appropriate,' or 'to do anything in line with laws.' So it didn't actually concede in any meaningful way."
His evidence: Pentagon spokesman Sean Parnell's tweet the day before, restating the official position — "We only allow all lawful use." Same words that showed up in the terms sent to Anthropic.
And "lawful" is precisely the word the entire argument is trying to get past. The whole concern is uses that are lawful today because the law hasn't caught up. Promising "lawful" as a guarantee against "lawful but harmful" promises nothing.
"The whole timeline has been driven by the Department of War, not by us."
My take: In any industry, "if the counterparty deems it appropriate" means the clause does not exist. But the tell is the tempo: three days, public pressure by tweet, a deadline set to the minute. That's not how you negotiate a deal. That's how you procure a refusal.
6. The President called him selfish. He answered with an offboarding plan.
Trump posted that day that Anthropic's "selfishness is putting American lives at risk, our troops in danger, and our national security in jeopardy."
Amodei's reply isn't on the same frequency:
"Even if they take these extreme actions, we'll do everything we can to support the Department of War to provide its technology for as long as it takes to off-board us and on-board a competitor who's willing to do these things that we are not willing to do."
He says the service interruption is what actually worries him:
"I've talked to people on the ground, uniformed military officers, who say, 'This is essential. Not having this will set us back six months, 12 months, maybe longer.'"
My take: Getting called selfish by the President and responding with "we'll keep the lights on until your replacement is running" is the strongest posture available here. It doesn't argue. It removes the premise of the accusation — that he doesn't care about the troops. Insults are cheap. A continuity commitment is expensive.
7. "Unprecedented": a label previously reserved for Russians and Chinese suppliers
"To our knowledge, the supply chain designation has never been applied to an American company. It has only been applied to adversaries — like Kaspersky Labs, which is a Russian cybersecurity company suspected of ties to the Russian government. Chinese chip suppliers. Being lumped in with them, it feels very punitive and inappropriate, given the amount that we've done for U.S. national security."
The bigger problem is how the designation spills outward. It isn't just "the Pentagon won't buy from you." It reaches into the private sector: if you hold military contracts, you can't use Anthropic in the work touching those contracts.
"So they're reaching in to the behavior of private enterprise. And it's very hard to interpret this in any way other than punitive."
Pressed on whether this is an abuse of power, he dodges once, then lands on a specific phrase:
"It was made very clear in some of their statements, in some of their language that this was retaliatory and punitive. I don't know what else to call it. Retaliatory and punitive."
My take: He could have said "yes, abuse of power." He chose "retaliatory and punitive" — courtroom vocabulary, not Twitter vocabulary. The word choice is the forecast. They're building a filing.
8. Why the Boeing analogy doesn't hold
Kent puts the obvious objection to him: Boeing builds aircraft for the military and doesn't tell the military how to use them. What makes you different?
"AI is moving so fast, I've talked often about how AI is on an exponential trend. The amount of computation that goes into the models doubles every four months. We have never seen anything like this pace of innovation."
The argument has two layers. First, aircraft are mature technology — a general has a pretty good understanding of how one works. Models aren't, and he doesn't. Second, and more fundamentally: Anthropic isn't shipping a component with a fixed spec sheet.
"Our model has a personality. It's capable of certain things. It's able to do certain things reliably. It's able to not do certain things reliably. And I think we are a good judge of what our models can do reliably and what they cannot do reliably."
My take: The most technical line in the interview and the most underrated. Boeing delivers a specification. Anthropic delivers a system whose behavior at the edges isn't fully predictable — and the vendor is the only party who knows where those edges are. That's where "you build it, we decide how to use it" starts to crack. This isn't an Anthropic quirk. It's a property of the whole generation of technology.
9. "If adversaries commit war crimes, shouldn't we commit war crimes as well?"
Kent keeps pushing the same direction: adversaries are catching up or have caught up. Isn't principle a luxury right now?
Amodei separates the two lines and prices them individually. Domestic mass surveillance — "does not help the U.S. catch up with its adversaries. It's an abuse of the government's authority, even where it's technically legal." Ruled out. Fully autonomous weapons — he concedes we may need to keep up, and the blocker is reliability and oversight, not principle.
Then he delivers the sharpest line of the interview:
"If adversaries commit war crimes, shouldn't we commit war crimes as well? I'm not saying this amounts to war crimes. What I'm saying is that the essence of our values is that we have to find a way to win in a way that preserves those values."
"We can't just be a total race to the bottom."
My take: The line works because of who says it. Amodei isn't a pacifist — he says nobody has emphasized the existential need to defeat our adversaries more than he has. Because he's unambiguous about having to win, the question of how doesn't read as posturing. Defend against them without becoming them: that argument only carries weight coming from a hawk.
10. Congress should own this. In the meantime, somebody has to draw the line.
Amodei concedes the arrangement isn't sustainable.
"I actually agree with you that this is not tenable in the long term. I don't think the right long-term solution is for a private company and the Pentagon to argue about this. I think Congress needs to act here."
The next sentence is his actual situation:
"Congress doesn't move fast. So I think in the meantime, we do need to draw a line in the sand."
On how it ends, he doesn't perform optimism. As of the interview, no formal action had arrived — only posts.
"All we've seen are tweets from the president and tweets from Secretary Hegseth. When we receive some kind of formal action, we will look at it, we will understand it, and we will challenge it in court."
He also points out that Hegseth's tweet overstated the legal reach — the law only bars contractors from using Anthropic as part of their military contracts, not from doing business with Anthropic at all.
"The nature of the tweet that the secretary put out was designed to create uncertainty, was designed to create fear, uncertainty, and doubt. But we won't let that succeed. We will be fine."
My take: "Draw a line in the sand" is the load-bearing phrase. He knows the position is temporary — a private company has no legitimacy to make this tradeoff for a democracy, and he says so first. What he's claiming is a transitional duty: until the rules catch up, the people who can see the problem first don't get to pretend they can't. Whether that argument holds depends entirely on whether Congress ever shows up.
Postscript: what else happened that night
The same evening Hegseth issued the ban, Sam Altman posted that OpenAI had reached an agreement with the Department of War to deploy its models on the classified network. Two of OpenAI's most important safety principles, he wrote, are prohibitions on domestic mass surveillance and human responsibility for the use of force, including autonomous weapon systems — and the DoW agrees with these principles, reflects them in law and policy, and put them into the agreement. He asked the Pentagon to offer the same terms to every AI company.
The same two principles. One company signed. One didn't.
The difference may live entirely inside phrases like "if the Pentagon deems it appropriate." Or it may not. We'll find out in a courtroom, or in Congress, or not at all.
Until then, one thing is worth holding onto: this round was never about whether AI goes rogue. It was about who gets to say no.
Sources: CBS News full transcript of its exclusive interview with Dario Amodei (February 28, 2026); CBS News reporting on the supply chain risk designation.