In a CNBC interview, Department of Defense Undersecretary for Research and Engineering Emil Michael discusses the Pentagon’s contentious negotiations with Anthropic over the deployment of its frontier AI models. Michael recounts a “whoa moment” after the Venezuelan raid that captured Maduro, when an Anthropic executive contacted Palantir— the prime contractor— to inquire whether Palantir’s software had been used, raising concerns about a subcontractor’s ability to impose policy constraints on classified operations. Michael explains that Anthropic’s demands— prohibiting mass surveillance of Americans and insisting on human‑in‑the‑loop control for lethal weapon decisions— conflicted with the Department’s existing 25‑page contract restrictions, making the negotiations untenable for a 3‑million‑person organization. He characterizes the mass‑surveillance demand as a marketing ploy and emphasizes that the Department already complies with the National Security Act, FISA, and other laws. Michael also compares Anthropic’s terms with those offered to OpenAI, noting that the government applied uniform contract language across frontier labs, and that any perceived differences were media‑driven. Finally, he addresses concerns that labeling Anthropic a supply‑chain threat could hinder the U.S. AI arms race against China, arguing that the decision was driven by operational risk rather than competitive advantage.
Chapters
Our next guest is at the epicenter of the Pentagon's feud with Anthropic. Joining us right now is Emil Michael. He is Department of Defense Undersecretary of Research and Engineering, and he heads up the Pentagon's adoption of AI. He also led discussions with Anthropic and its CEO, Dario Amodei, on the use of the company's AI models. And first of all, Undersecretary Michael, thank you for being with us today.
Good to see you, Becky.
Good to see you too. Let's walk back and try and figure out what happened. We've spent an awful lot of time talking about the government's battle with Anthropic and kind of the fallout from it. But if I understand correctly, I think what first happened was a holy cow moment. I'll say that you had. I think you called it something slightly different, but a moment after the Venezuelan raids that captured Maduro. Where Anthropic went and asked Palantir if its software had been used in those attacks. And you said at the time that that was your whoa moment. What happened?
Yeah, what happened is we're in the middle of negotiations with all the frontier lab companies to more widely deploy their models across the department for all the various use cases— enterprise intelligence, warfighting. And then, after the Maduro raid, one of the senior executives from Anthropic had called Palantir, who is the prime contractor and Anthropic is a subcontractor in a contract with the Department of War, and asked if their software was used, which is classified information, and implying that if they didn't like the way it was used in that raid. That we might have a problem and have violated the terms of the contract, which meant that they could shut us off at any time, which was the whoa moment, because then we realized we are dependent on this one provider who wants to insert their policy preferences in the middle of an operation, potentially, and harm the warfighter. And that was something that we took extremely seriously at the time.
Did Palantir answer the question that was asked? Was their software used?
I can't say. It's not okay for me to say here. Yeah, it's still classified. But the mere asking of the question and then the mere relaying of that question to other executives at the company, who have since commented that they were asking different kinds of questions and they didn't have clearances, really posed a big problem. But really, the substance of problems there was that they were worried about their software being used in a raid. And remember, they've sold their product into what's called a combatant command, that engages in combat. And their terms of service, you know, try to restrict how that combat is conducted. And that means they're inserting themselves in the chain of command. And that was very worrisome to us at the Pentagon.
Okay, let's walk through what we on the outside have kind of heard and how it's been reported and how we understand it. The sticking point came down to 2 points in the negotiations that you were leading with Anthropic, and that was, first of all, they didn't want their, their software to ever be used for spying in mass levels on Americans, whether it was legal or not. And second of all, They didn't want decisions being made without a human involved when it came to destruction and weapons that could kill— that could kill people. Is that the case?
No, it's sort of a red herring. When we got the contract that we inherited from the Biden administration, there were 25 pages of different restrictions on the Department of War and its use cases. And they were so restrictive when I looked at them that we started negotiations to sort of reset The baseline. I mean, we are the Department of War. We use their software and services and hardware for Department of War activities, and we wanted to baseline where we're at. And so, at the end of it, we actually got to terms on the autonomous question because we have our own policies at the DOW which say that there has to be human oversight on autonomous weapons, and there has to be human oversight on the development of autonomous weapons. Which you might use physics, material science, aerodynamics, all those kinds of things. But throughout the negotiation, we had to negotiate every single exception to their broad use policy. And that became untenable in a department of 3 million people. The mass surveillance thing was just a marketing ploy, really. The Department of War has no equities in that kind of thing. We follow the law. We actually agreed to put in the contract— That we follow the 1947 National Security Act, the FISA Act, and all other applicable laws. They were trying to redefine their own definition of what domestic surveillance meant and take that power away from Congress and our democratically elected leadership.
Emil, let me ask what happened with OpenAI, because you were negotiating with OpenAI at the same time. And from the outside, it looks like OpenAI received basically the same considerations that Anthropic had been asking for on those 2 points. Is that the case? Why did you strike a deal with OpenAI and not Anthropic?
There was a few points here. So we told every frontier model company that every, every lab gets the same terms because we're the government. We have to be fair. And frankly, we wanted all 4 labs to have their products available Right. —to our warfighters. And so, the contracts that we wrote are very similar in what they provide. And different types of language got people more comfortable than others. So, there was really no big difference between the two. Remember, at the Department of War, we have 2 types of agencies. We have the combatant commands and the Department of War, where the civilians work. We also have intelligence agencies, and we're knocking them down one by one. So OpenAI got the same terms on combatant commands. We got the same terms on the rest of the Department of War. And we were just trying to finish up the intelligence agencies there. So, you know, another obfuscation that Anthropic put into the, into the media space for some reason.
From some of the conversations we've had around this table, is that we are in an arms race right now with, with China in particular. We believe we're at the front with, with AI. Anthropic clearly is one of the leading firms in that fight. By going a step further, obviously the government can do business with whoever it wants, but by taking that further step and declaring them a threat to the supply chain, meaning that other contractors to the government can't use them, that probably puts Anthropic in a pretty perilous position. In terms of continuing to, you know, take on the Chinese and be able to develop at a faster pace. What, what do you say to that argument?
Yeah, I'd say that the way their executives conducted themselves by asking for classified information and communicating messages that should be classified among their executives, bad faith negotiations, and then The other really subtle point that I think we're going to hear a lot about AI in the coming year: insider threats, model poisoning. Remember, their their model has a soul, has a constitution. That's not the U.S. Constitution. The other day, their model was anxious, and they believe it has a 20% chance right now of being sentient and have its own ability to make decisions. So do does the Department of War want something like that? In their supply chain so that it could hallucinate, it could corrupt models that are used by defense contractors who are building weapons systems or airplanes and so on. So the truth of it is we can't have a company that has a different policy preference that is baked into the model through its constitution, its soul, its policy preferences pollute the supply chain so our warfighters are getting ineffective weapons, ineffective body armor, ineffective protection. And that's really where the supply chain risk designation came from.
