The Pentagon's AI ultimatum.
Ben Yelin: Hello, and welcome to "Caveat", N2K CyberWire's privacy, surveillance, law, and policy podcast. I'm Ben Yelin from the University of Maryland Center for Cyber Health and Hazard Strategies. Dave is out today, so I'm joined by my N2K CyberWire colleague, Ethan Cook. Hello, Ethan.
Ethan Cook: Hey, Ben.
Ben Yelin: On today's show, Ethan and I have the story of the Pentagon's deal with Anthropic falling apart in dramatic fashion. While this show covers legal topics, and I am a lawyer, the views expressed do not constitute legal advice. For official legal advice on any of the topics we cover, please contact your attorney. [ Music ] All right, Ethan. It's the two of us again. We are without our fearless co-host, Dave Bittner, but it is good to have you with us.
Ethan Cook: Hostile takeover.
Ben Yelin: Yes, we are. This is a hostile takeover. No, Dave is, I'm sure, doing much more important things somewhere. But we are lucky in that we have a very exciting story to cover, and this is about what happened with the Pentagon and their contract with AI startup Anthropic. I don't know if you want to start by just kind of giving us a basic overview of what happened, and then we can get into the details.
Ethan Cook: Yeah, so obviously, the Pentagon, as with many of the other government branches, contracts out with multiple AI providers, OpenAI, Anthropic, et cetera, using their AI services. And from what I understand, this all ties back to when the U.S. abducted Maduro from Venezuela. And there were some internal rumblings, this wasn't public yet, between Anthropic and the Pentagon, that Anthropic was not okay with some of the things that were being requested of it. Specifically, the DOD's plan to use its AI services for both autonomous weapon systems as well as mass surveillance of U.S. citizens. And they wanted very strict stipulations put in that their services were not going to be used for that. And apparently, they were used in a way for the U.S. incursion into Venezuela, and that started this feather rustling. And from there, things kind of continued to get a little bit more stressful.
Ben Yelin: Yeah, so going back a little bit, Anthropic signed a deal with the Pentagon, a contract back in the summer of last year. It was worth up to $200 million, and it was going to be the first AI lab to deploy its models on the Pentagon's classified system. So this was a big deal at the time, got a lot of news coverage. And then what you're talking about kind of was a murmur happening below the radar, kind of behind the scenes. Is that right?
Ethan Cook: Yeah, so they had been working together. They had been building out this relationship. And this incident happened. And, again, this was not public yet. There was some tension building. And then suddenly, about a week ago, it really started spilling over into the public sphere with reports that the Pentagon was going to drop them and that Anthropic was not budging. And the Pentagon side of this story was that we asked them this hypothetical question whether they would let us use their services for defense, for missile defense programs to help. Like if someone launched a nuke at us or an ICBM, that they could help, we could use their AI services to help coordinate a defense response effort, et cetera. Which Anthropic -- the Pentagon characterizes their response as, oh, yeah, just give us a call and we might help kind of vibe. Like, you know, it was like, oh, we'll see what happens. And Anthropic did not like that characterization one bit.
Ben Yelin: Yeah. So as with most things that happen these days, the major news came out in a post on Truth Social from the president of the United States.
Ethan Cook: That's where the best news comes from.
Ben Yelin: It is. So that post said that Anthropic was a woke company and was going to destroy -- I'm paraphrasing here -- but was going to destroy the warfighting powers of our Pentagon. And therefore, we are severing ties from Anthropic. And not only that, all federal agencies, not just the Pentagon, are prohibited from using Anthropic's technology. Anthropic, in his view, was trying to strongarm the Pentagon. He said that he would not be afraid to impose civil and criminal penalties if the company did not cooperate during their six-month phase-out period. So that's a key element of this. And then furthermore, later that same day, defense secretary/fitness expert Pete Hegseth formally designated Anthropic as a supply chain risk to national security. So usually, if something says supply chain risk to national security, it's going to be like a Russian or Chinese company that's a cog in our supply chain where we're concerned for national security purposes that they might do something to our products or services. So it's been used, for example, to restrict the company Huawei, which I know we've talked about on this podcast. It has not been used for an American-based company like Anthropic. So as a result of this designation, Pentagon contractors, of which there are a lot of -- I'm sure you've been to Northern Virginia. It's beautiful.
Ethan Cook: Yeah. Contractor City.
Ben Yelin: Contractor City, yeah. They all have very nice offices on the way to Dulles Airport. And we are migrating away from the AI service Claude, which is Anthropic's baby, across our ecosystem. Anthropic was not happy about it. They challenged this designation in court. And we'll see what happens there. But then the Pentagon very quickly found their replacement. Do you want to give us some information about how that happened?
Ethan Cook: Yeah, so like you mentioned, Pentagon, after they tried to come to some deliberations, I think Hegseth put in a deadline. It was like by the end of the day Friday, if Anthropic didn't recant its decision and go along with them, they were dropping them and labeled them a supply chain risk. Anthropic said, we're not budging. Pentagon subsequently dropped them. And the very next day, OpenAI gets a little ring. And OpenAI is now --
Ben Yelin: Hello, this is Sam Altman here.
Ethan Cook: Exactly. Exactly. And they're requested to basically fulfill that hole that the government has just made and in walks the next person. It felt like when I was reading these stories, by the way, I was watching an episode of "Real Housewives" or some reality TV show. Where it's, you know, I don't like you, I'm going to this person, you know, swapping, swapping, swapping, who can offer me the most? It felt very dramatic. The interesting thing that kind of came out to me was not just the swap partnership. You know, I expected a swap if they're going to drop Anthropic. I expected them to switch somewhere. But the thing that I thought was most interesting was the sworn statements that they would, from the U.S. government, that they would not use their OpenAI systems, ChatGPT, et cetera, for any mass surveillance or any weapons of automated weapons. Which that kind of then brings this kind of question into back to the Anthropic side of things, which is, okay, if they're willing to sign that for OpenAI, then is that kind of a shallow thing? What is going on here? Because it seems like everyone's contradicting each other.
Ben Yelin: Yeah. And it just raises a bunch of questions like what is different about this agreement that made it acceptable to OpenAI and to the Pentagon? So there are a couple of really interesting elements to it. For one, and I am not an expert on this, and you probably know more, like, I think at this point in their evolution, Claude is a better AI model than ChatGPT.
Ethan Cook: Yes. Yeah, I would say so.
Ben Yelin: Now, I don't pay for the good version of either of those.
Ethan Cook: Neither do I.
Ben Yelin: If any sponsors want to help us out with that, we can give you firsthand knowledge. So that's kind of one element of it. But then this question of, like, what exactly did OpenAI agree to, and why potentially should we be suspicious? So from what OpenAI said is they guaranteed safeguards to protect against domestic surveillance and to protect against the use of autonomous weapon systems, which, if that sounds very familiar, it is. If you read between the lines, the difference is that OpenAI is accepting a legal framework that references existing laws and Pentagon policies. They're not looking for these contractual assurances that Anthropic was looking for, which says, like, you need to come to us if you are intending to do these things. And we are going to be very protective of our models and institute our own human controls. OpenAI is just like, we promise we won't break any laws. Here's ChatGPT. Enjoy. And I think that level of cooperation is what the Pentagon was looking for. But the fact that this happened literally within hours of the Anthropic deal falling through, I can understand why people are suspicious about exactly what OpenAI may have agreed to.
Ethan Cook: Yeah, and I think it just also raises a stability concern for me, where, you know, we are, you know, back in the summer -- taking a step back, Trump released his AI action plan in the summer of 2025, which a key portion of that was investing in AI companies, American AI companies to boost innovation, secure defense, involve them in defense projects, et cetera. Right? And Anthropic, as we have just said, is one of the top three and arguably the best AI developer in the States. And for us to alienate them and, you know, almost create a semi single point of failure by going after OpenAI exclusively, I would have loved to see a dynamic where we use both. Right?
Ben Yelin: Right. Take advantage of American ingenuity.
Ethan Cook: Exactly. And also make the two compete against each other. There's this dynamic where, you know -- and who knows, maybe that this whole relationship becomes positive again and in six months, Anthropic's back into where they were or whatever happens. That would be great. But if it doesn't, I kind of sit here and walk away going, the only group this really hurts is the U.S. as a whole. It doesn't help competition. It doesn't advance innovation. It creates some really ethically gray zones with how the defense department can create use AI because we're not really solving the problem. And it seems a little bit more like vibe law, to borrow from vibe coding. It's like, well, maybe it's illegal rather than really kind of honing and ironing out these questions and these relationships. And I sit here and I wonder and I'm curious if you have any similar thoughts of where does this like leave other AI companies who aren't OpenAI and aren't Anthropic?
Ben Yelin: I mean, I think it puts the entire industry in a really difficult place. You are torn between wanting to win big government contracts, which all of them want to win, and your own stated ethical principles. And companies have tried to toe this line forever. And generally, you know, there have been some hiccups here and there, but it's worked out. I mean, government agencies are able to come up with agreements that ensure the ethical use of these products. Here, though, we're talking about two very serious things, mass domestic surveillance. In case you were not conscious during the Edward Snowden exposure era, that's something that's quite controversial in this country.
Ethan Cook: Yeah.
Ben Yelin: And then using autonomous weapons systems --
Ethan Cook: Like that also would never --
Ben Yelin: Yeah, I mean, and that's like very, very serious business. So you saw that there was a bit of a revolt in the industry across the AI sector, so within these big companies. You saw 60 OpenAI employees who voluntarily resigned because they backed Anthropic's original stance. You saw employees from other companies writing open letters, standing up for Anthropic and expressing their dismay at what happened with OpenAI. So you also might be creating this divide within the industry between the people who write the contracts and the developers, the people who are building these products who don't want to see the fruits of their work used to autonomously kill people in foreign conflicts or for mass domestic surveillance.
Ethan Cook: Yeah.
Ben Yelin: And if you see a revolt on a large scale, then I think that could be extremely problematic for the industry.
Ethan Cook: Yeah, it feels like this is almost a race to the bottom that like now that OpenAI has opened this door, right? It's like, okay, who's willing to cut corners and bend to military demands the most to get the contract rather than having that established set of principles of we won't engage in surveillance, like we're just not going to. Or if it's surveillance, it has to be Congress approved and, you know, legislatively, it's going through this whole process. And it feels like because there's such a lack of -- what's the right word for it? Maybe lack of existing laws having AI stipulations built into them. It feels like, you know, this is a prime time for someone who wants to take advantage of that lack of framework to do so. And that feels like what the guy -- and I wouldn't be surprised if the military has done that. I mean, like, you know, I'm not this is not like a pro or against, but the military takes the advantages that it's given. And if this feels like a perfect opportunity to use OpenAI for whatever, any AI system for that, because they can walk away saying look, there wasn't a law that prevents us. There was nothing that said that we couldn't use this. We get nothing illegal, you know, kind of a free blank check, so to speak.
Ben Yelin: Yeah, and they're doing that under the color of the law, basically saying, like, we agree only to provide services in a lawful manner, consistent with the Pentagon's current laws and policies. That gets problematic for a couple of reasons. One, I mean, do we actually trust them considering that they made this agreement to really push back if they're given some type of illegal order, given that they were able to come to this deal within hours of the Anthropic deal falling through? And then also, there are just a lot of gray areas, as you said. Like, this is an area of the law that's just not very developed. So they might say now that we're going to comply with all laws, but there are a lot of edge cases. And if their main goal is maintaining the status of their very lucrative contract, that creates a pretty perverse incentive structure, in my opinion.
Ethan Cook: It sends a message to -- you know, and if I'm happy to see people are resigning and standing up because, like, to me, I sit there from the AI community. And if something has to be done coming from internal rather than external, it needs to almost rebuke this because the law is, I'm sure you're aware, you know, Ben, in your days working both state and federal, it moves a little slow sometimes.
Ben Yelin: You don't say.
Ethan Cook: It's not the most updated. We've had AI for years now, and we still have no federal AI framework from a legislative standpoint. So, you know, we have these massive holes that aren't anywhere close to getting formed and probably won't get formed for the next at least two years. And this is the time where if you're going to kind of walk that gray zone where you're like, oh, well, I'm technically not doing anything illegal. I feel like if you have to caveat with the technically, it's really concerning. And I feel like that's where we're heading to right now with this OpenAI deal. And I hope that the rest of the AI community doesn't really engage in this dynamic.
Ben Yelin: All right. We're going to take a quick break. More to come after this. [ Music ] We oftentimes have been told that, like, one of the ways we can fight back against surveillance or any other violations of civil liberties is to vote with our wallets. So if Apple has great security protocols and end-to-end encryption and it promises to never divulge information except under a lawful warrant, and it fights back when the FBI tries to obtain data about the iPhone or the San Bernardino shooter, well, if we care about that stuff, then we will purchase Apple products. That system falls apart when you only have a few AI companies. And again, the big players, there's probably at maximum five of them. So it's not a duopoly necessarily, but there's certainly consolidation within the industry.
Ethan Cook: Yeah.
Ben Yelin: And I'm sure most of the startups are going to be bought out at some point by one of these big companies.
Ethan Cook: Yeah, if not collapse.
Ben Yelin: Or collapse under its own way, yeah. So there's just not a level of competition. And then when you have this race to the bottom, where's the assurance to the consumer that they are competing for your services as a privacy-conscious consumer and not for lucrative government contracts?
Ethan Cook: And I think to echo that, there's a dynamic where these companies actually don't even care about making a profit. Almost all of them run a negative, right? Like, they're all -- so to me, they're not even -- like, obviously they're profit-oriented. They want to make money, but they care far less about the day-to-day making money than more so going to an investment round and raising $500 billion. Right?
Ben Yelin: Right. It's the VC meetings that they are most concerned about.
Ethan Cook: Exactly.
Ben Yelin: Yeah, and so if you can go to the VC meeting and be like, hey, guys, I got an in with the Pentagon. We have knowledge that the Anthropic deal is falling through. And if you invest in us now, we'll be right there to take on this ginormous contract. And you guys will be able to afford your second and third private jets. Not to sound too cynical, but it's just, I think you capture the dynamic well of what's actually happening out there. And I think it doesn't bode well for a normal consumer that wants to use these tools, but at this point is unable to trust the companies to put ethics at the forefront.
Ethan Cook: Well, to the point, right, OpenAI just closed their most recent fundraising round, like, less than a week ago. $110 billion, of which the big, large contributors were Amazon, NVIDIA, and SoftBank. And so I am sure that the concepts of what are -- I don't know if the specific deal themselves came through for the Pentagon, but I'm sure that there were conversations of, are you trying to break into the government sphere? If you do, what is the money that we're in? These are massive figures of numbers being thrown around. This isn't something like, hey, oh yeah, maybe we'll go into the Pentagon. Maybe we'll go into DoD. You know, that's a conversation that is absolutely coming up, right? Especially as this fundraising round is closing and being announced, right as this other deal is being announced in the process. It's just, you know, I feel like a little bit of a conspiracy theorist thing. Like, oh, it's all convenient timing, but it's, you know.
Ben Yelin: It's all convenient timing. Yeah.
Ethan Cook: Exactly. You know, it's, there's a, one company falls, the other company gets the same deal, and then it inks another $110 billion worth of investment. Right? And for Anthropic, like the $200 million, when you're talking $110 billion for fundraising, $200 million is not a lot, right? That's chump change. But it's the premise of, okay, we're starting here, and then who knows where it goes from there? Like $200 million is the dipping the toe in the water, kind of, so to speak.
Ben Yelin: Right, right. Not only that, but then when you have this policy that bans all contractors from using Anthropic -- and we'll see what ends up happening with that. I mean, it's kind of unclear at this point if they're going to come up with some deal to resurrect the product for contractors. But that's where we stand as of this recording. But then, you know, that threat makes it even more serious. So if you are an enterprising company, and you're in this sphere, and you want to hold yourself out to your consumers as being conscious about things like mass surveillance and autonomous weapon systems, now you see that if you try to stand up for those principles, you not only lose your $200 million Pentagon contract, which again, maybe, as you said, you can do without that. But you also lose access to military contractors, and these contractors are huge entities. We're talking Palantir here. We're not talking about, like, mom-and-pop shops who have a sub-sub-sub contract with the Pentagon. Like, these are some of the biggest firms in the country. So if you're blacklisted by the government, I mean, that could be a fatal blow to your service. And what does that mean for these other companies who want to get into this space? I think perhaps there's somewhere in the market for a company to say, we will never sell our services to any government agency. We will be the ethically conscious, privacy-focused AI tool, and if you are an AI user and you want to purchase a product, a premium service, do it through us. And maybe that's the answer here. The problem is that the administration in this and other spheres have reacted to that type of threat with threats of their own, or, you know, they'll be blacklisted, or there will be some type of investigation. So we're just not in a place where the market can be supportive of people who have that ethics or privacy focus.
Ethan Cook: Yeah, I think there is a dynamic here, you know, especially given how hawkish this administration has been. I mean, it's been a year, essentially, and we've, you know, we still have Ukraine and Russia going on. We have Israel-Palestine. We've gone after Venezuela, and we've gone after Iran multiple times now. To me, the role that the military is playing is incredibly aggressive. And with this administration's very, very clear backing of AI and AI use in innovation and in defense, to me, like, if it was a different administration -- again, this is not me trying to critique and be like, oh, I disagree with them using them in defense. I would sit there and be like, okay, maybe there's a different perspective here, but I don't see a world in the next three years where the administration doesn't look at OpenAI or name the other AI companies that they contract with and say, how can we use your services to advance our operations and skirt that ethically gray zone because we have no law in place? It's just, that feels like something that is going to come up in the next three years at some point if it's not already coming up with OpenAI and how we are currently engaging in Iran.
Ben Yelin: Yeah, I mean, and that's another sort of, I guess, again, not really a coincidence in all of this, is this happened on Friday, and we all woke up Saturday morning to news that we had started an air invasion along with Israel of Iran. And so, I think the timing here is absolutely critical. We are going to start deploying these tools in armed conflict.
Ethan Cook: Yeah.
Ben Yelin: Which makes it, to me, even more puzzling that OpenAI, which probably knew that this was on the horizon, maybe they didn't know how quickly it was going to happen, but they knew that.
Ethan Cook: Yeah, there's no way that they would be blindsided by it.
Ben Yelin: Exactly. And so the fact that they at least should have been aware that that was a threat, and they still, within hours, agreed to this deal that's not based on any sort of ethical framework, but it's based on a, we promise to comply with all the laws, pinky swear agreement, I think that's certainly concerning. And I'm wondering if there's enough knowledge among consumers that this will have any impact on OpenAI's bottom line. I tend to think it will not.
Ethan Cook: I would agree.
Ben Yelin: Yeah, I'm wondering what your perspective is on that.
Ethan Cook: I don't think it will. Well, I think, you know, one, I think OpenAI is probably hedging the bet here that the long-term gain of getting the DoD early is worth any PR hit, you know, or any impact on consumer vision of them. I think, two, because we already mentioned that they, right now, at least seem to be the, you know, they're losing the race to Anthropic, quote-unquote, that, okay, time to play a little riskier. You know, we can't sit on the high horse and be like, okay, we're, you know, we're winning this race. You know, we kind of have to do things that we wouldn't have done a year ago. And I think they're willing to take that bet. With this administration being so very pro-AI, I think there is an implication saying here, okay, whatever negative hit we take, getting into the graces of the Trump administration, even if it's through the DoD, if they like us because we did that, that may open the door to other AI projects that are not related to defense at all and give us opportunities to form deals, give us opportunities to expand data centers, et cetera, or cut other hurdles or influence legislation that we want to get things done faster, et cetera, whatever it may be, that they probably wouldn't have had otherwise. It's almost like a you scratch my back, I scratch yours kind of quid pro quo here.
Ben Yelin: Yeah, I've just kind of, as a closing thing here, I'm reading the Our Agreement with the Department of War blog posts on OpenAI's website.
Ethan Cook: I have not read it yet. I'm sure it's a fascinating one.
Ben Yelin: It is. I mean, it seems to me like gaslighting, like it's kind of insisting that they have agreed to these kind of the same protocols as Anthropic, maybe even more so, that they're more protective against domestic surveillance and use of autonomous weapons. And for things like use of OpenAI technology for high-stakes automated decisions like social credit, they're very adamant about that. It's a very strongly worded post here. But it's just one of those show me through your actions and not with your words. And I know you have to make a logical leap here, but if Anthropic was completely blacklisted in the federal government for fighting for these things, and OpenAI wins the contract and is not blacklisted, does that not imply that they have not secured any type of robust guarantees?
Ethan Cook: That's what I was saying earlier, which was confusing to me, right? Like, Anthropic comes out saying, we're not working with them because they want us to do mass surveillance and autonomous weapon systems, and the DoD says, no, that's not what happened, right? DoD says it's something else. The two guys don't disagree on why this relationship fell out. And then suddenly a new OpenAI deal is inked, and then in that deal, the deal says, hey, we will not use the weapons for our AI systems for our autonomous weapon systems and mass surveillance. So I sit there and go, okay, wait, if that was the fallout point, then why is it not the fallout point here? What is the difference in these relationships? Like, what has fundamentally changed between these two and who's willing to accept what? Because it sounds outlandish to me that Anthropic would say, no, you can't use our AI systems for missile defense, right? Like, it's like shooting down a nuke, right? Like, I think anyone would want that to happen. If California gets nuked, Anthropic goes with it.
Ben Yelin: Right.
Ethan Cook: So, like, I don't know why they would disagree with that. So -- what?
Ben Yelin: It's very bizarre to me. And I think kind of looking at the future here, I'm curious as to what impact this has on contractors. We'll hear what contractors have to say about their forced migration away from these systems. I'm interested in monitoring the migration process to see if the administration holds fast to the timeline. And also, if they kind of TACO out, Trump always chickens out. And at some point under the radar in three months, it's like, yeah, actually, it's fine to use certain Anthropic tools. So I'm kind of curious about that. And then just what the fallout is going to be within the industry among the rank and file who have built these tools and now see a company that's seemingly, even though they're adamantly denying it, giving in to these pretty extreme government demands. And what that's going to do to the workforce. So we have a lot to look forward to in this story.
Ethan Cook: Yeah, I think this is the initial dip the toe in the water. And it's going to, over the next six months to a year, really, really start heating up. I wouldn't be surprised if Congress gets involved and starts asking some questions. I wouldn't be surprised if it gets messier publicly or if more news comes out. I'm really curious to see what Anthropic does from a PR standpoint over the next month and what they try to characterize this as because I don't think we're done with this.
Ben Yelin: And I think I'm also curious to see what happens with the legal process, too. Is the sort of supply chain risk restrictions that they're instituting here, is that legal when you're just trying to punish the company for how it's engaged in negotiations or what it's demanded in negotiations? Or do you actually have to have some prima facia showing of a threat to the supply chain? If this is just punishment for punishment's sake, then I think courts up to and including the Supreme Court are going to have problems with it. But if you can have an actually persuasive justification that this is a national security risk, which we'll see, I'm extremely dubious about that. But that's something I'll be looking out for as well.
Ethan Cook: Yeah, I'm curious to see how that plays itself out because I think right now you could argue all AI is a national security risk.
Ben Yelin: Right, yeah.
Ethan Cook: And not because of anything that Anthropic does, just because I don't think these systems are inherently secure and I don't think they're inherently focused. So I would be curious how they approach the argument of, is it specifically Anthropic or are they going to approach it as, oh, AI in general is not secure. So we just decided that this company was less secure than the other ones and that the other ones were taking more steps in how they characterize this supply chain. I think it will be a very interesting trial.
Ben Yelin: For sure. Well, Ethan Cook, my colleague at N2K CyberWire, thank you so much for joining us and filling in in this episode as a worthy replacement for the incomparable Dave Bittner.
Ethan Cook: Well, thank you for having me. I enjoyed the conversation.
Ben Yelin: And that's "Caveat", brought to you by N2K CyberWire. We'd love to know what you think of this podcast. Your feedback ensures we deliver the insights that keep you a step ahead in the rapidly changing world of cybersecurity. If you like the show, please share a rating and review in your podcast app. Also, please fill out the survey in the show notes or send an email to caveat@N2K.com. This episode is produced by Liz Stokes. Our executive producer is Jennifer Eiben. The show is mixed by Trey Hester. Peter Kilpe is our publisher. I'm Ben Yelin.
Ethan Cook: And I'm Ethan Cook.
Ben Yelin: Thanks for listening.
