← The AI Policy Podcast
The AI Policy Podcast
Anthropic Goes to Court While Claude Goes to War in Iran
01:03:17
In this episode, we explore the Anthropic-Pentagon clash following recent updates.
Loading summary
Transcript76 lines
- [00:00]
A
Foreign. Welcome back to the AI Policy Podcast. I'm Matt Mand and today Greg Allen and I will be covering all the latest updates in the Pentagon anthropic clash, as well as how AI is shaping the war in Iran. Greg, thanks for joining me.
- [00:22]
B
Great to be here.
- [00:23]
A
Well, we usually don't discuss the same topic two episodes in a row, but so much has happened in the Anthropic Pentagon saga since our last recording, so we had to Greg, you've been on a whirlwind media circuit doing interviews with Bloomberg, cnbc, abc, BBC, Fox News, Wall Street Journal, Washington Post, and a long form podcast interview with Ben Thompson. Usher Tekeri. So help our audience get up to date here. Where did we leave off in our last episode and what has happened since?
- [00:49]
B
Well, I think a ton has happened and one thing that's like important to know is like, even though there's a lot of information in the public domain, like not all of it is correct. There was a New York Times daily episode about this topic and most of it was good, but they got some like, important details wrong, such as saying that Anthropic's views towards both lethal autonomous weapons without human oversight and domestic surveillance are of the not never, but not now claim. And that is actually only true of Anthropic's position towards lethal autonomous weapons without human oversight. That is not characteristic of their opinion vis a vis mass domestic surveillance. In that regard, they never want to do it. It's not about the reliability of the technology. It's about their moral beliefs on the use of the technology, regardless of its reliability level. So there's a lot of inaccurate information out there to the extent that we can. We're going to try and provide the best sort of soup to nuts overview of where we've been, where we are, and where we're likely going in, which, as the episode title suggests, is two shockingly contradictory places because Anthropics Claude is going to court. But Anthropics Claude has gone to war in what is certainly the most significant military operations that the United States military has undertaken since the withdrawal from Afghanistan. It's just an incredible time to be alive and really shocking just for me in my career because I remember, you know, 10 years ago when I was saying AI and national security, it's going to be a big deal. And then you fast forward to today and like, it is a big deal. And we learned so much from these court filings. It's actually kind of shocking if you haven't read through them. But like the extent to which Anthropic discloses the use of its technology in military operations, just what it's being used for is pretty dang interesting. And there's some instances where you can kind of see that they're trying to dance around not disclosing classified information. So, for example, you know, they say that Anthropic is being very, is being used extensively for all categories of cyber operations. And then they say reportedly Anthropic is being deployed to Iran. So you can tell that classified information is really a challenge, and that actually matters in a big way for who's going to potentially prevail in this legal context. So we'll get into all that. One caveat that I do think we have to say up front is I'm going to do my best here to understand, you know, the nuances of these court filings and how they intersect with specific provisions in the United States law. But I'm not a lawyer.
- [03:56]
A
Yeah, neither am I.
- [03:57]
B
So even as I was just, you know, saying how you can't trust everything you hear in the news about these complicated issues and saying, you should listen to us instead of, now let me directly contradict myself, I'm going to give you my best understanding of the situation, which may be faulty in certain circumstances, but again, to the best of my understanding. So that's, that's that. Now, Matt, I believe you asked, like, what exactly has happened?
- [04:22]
A
Yeah, you know, where, where have we been? You said where we've been, where we are now and where we're going. So let's start with the first, the first one of those.
- [04:28]
B
Yeah. So our last episode on this topic was February 25th, and that was right after Secretary of War Pete Hegseth had presented Anthropic CEO Dario Amadei with the so called ultimatum. And that was where Axios reported that, quote, Hegseth told Amadei in a tense meeting on Tuesday that the Pentagon will either cut ties and declare Anthropic a supply chain risk or invoke the Defense Production act to force the company to tailor its models to the military's needs. Anthropic had until Friday, February 27th, to respond. Well, where are we now? Anthropic and the Pentagon were negotiating up until the exact moment of that designation as a supply chain risk up until the time of the ultimatum. Those negotiations are now over. In fact, there was even some negotiations that took place after Pete Hegseth's tweet on the topic. So Trump posted on Truth Social ordering all government agencies to stop using Anthropic products and is now reportedly, according to Axios again, which has been doing a lot of the best work on this topic, is prepared in executive order. On the same topic, Secretary of War Pete Hegseth has not just had this tweet, which he has framed as a final decision prohibiting Anthropic from being used across all. Basically, anybody who does business with the Department of War in any way cannot do business with Anthropic in any way. That's how the tweet is framed.
- [06:02]
A
Yeah.
- [06:03]
B
But then there's also been the formal letter sent to Dario and Anthropic formally designated a supply chain risk, which is not quite as broad as what's in the tweet. But from the perspective of Anthropic's court filings, the Trump Truth Social post, the Hegseth tweet, and the official supply chain risk designation all matter in a big way to the sort of legal case that Anthropic is bringing. So even though some actors in the ecosystem are ignoring the Truth Social post and ignoring Hegseth's tweet and focusing on this SCR designation, that's actually not the legal strategy that Anthropic has taken here. And we'll see. That's going to be really important. So these lawsuits, there's now one in the appeals court of the Northern District of California, and then there's actually just the Northern District of California and then United States Court of Appeals for the District of Columbia. They filed two separate lawsuits. There's some wonky reasons why they had to do that. It was clearly what they had to do in this regard. But the information in this document are pretty dang extraordinary. And so we're going to break them down.
- [07:12]
A
Yeah. So before we unpack those in depth, there's another part of this story that I feel like we can't ignore, and that's the OpenAI part. So the same day, I think it was, the February 27, that we saw this tweet from Trump and then the follow up tweet from Hegseth, we also saw an announcement from OpenAI that they had agreed to the Pentagon, with the Pentagon, to put their models on classified networks. So can you explain a little bit about how that unfolded? Because as if there wasn't enough drama already just with anthropic, OpenAI stepped in. And this is a whole other part of the story that I feel like is relevant here.
- [07:46]
B
Yeah. So if I had to guess, and this is me speculating, this is why a deal didn't happen. Because on March 4, the information leaked an internal memo sent by Dario Amadei criticizing OpenAI and the Trump administration. So right after Anthropic had failed to reach a deal and those very provocative social media posts came out by the administration, OpenAI announced that it had made a deal and claimed to have, you know, basically found the middle ground that Anthropic and the Department of War were unable to find, that addressed the safety and liberty and rights concerns, but also addressed the military necessity concerns. Well, we know what Dario thinks of that deal because his internal memo was leaked to the information. And I'm just going to read some of the choice quotes from that memo. Quote, the real reasons the Department of War and the Trump administration do not like us is that we haven't donated to Trump while OpenAI, Greg. Have donated a lot.
- [08:52]
A
It's not you. That's not you, Greg, right?
- [08:54]
B
Yeah, that's not me, Greg. That's Greg Brockman, the president of OpenAI. Quote, we haven't given dictator style praise to Trump while Sam has. Quote, we've actually held our red lines with integrity rather than colluding with them to produce safety theater for the benefit of employees. And this is a remarkable ding quote due to selection effects. They're as an OpenAI sort of a gullible bunch. But it seems important to push back on these narratives while which Sam is peddling to his employees. Now that's. I just want to double click on that, that burn that Dario just slung at OpenAI's employees when he says selection effects, he's basically saying, we, the seven members of Anthropic who left OpenAI to found Anthropic. We were like the non gullible bunch, the people who hadn't fallen for Sam Altman's propaganda. And ever since then, the people who are gullible go work for Sam. The people who are not gullible come work for us at Anthropic. I'm not claiming that that is truthful. I'm just saying that, wow, that is an incredibly vicious criticism. And so if you ask yourself, you know, what was underneath Sam Altman and Dario Amadei not holding hands on the stage at the India AI Summit, it is this kind of sort of deeply held resentments towards each other. So I believe that the leaking of this memo specifically torpedoed a deal that might have been reached. Even after the Trump post, even after the Hegseth post, Emil Michael, the undersecretary of war for Research Engineering, had said that negotiations were continuing and that he was open to negotiations. There was Some reporting that came out of the Financial Times and elsewhere that said that those negotiations were actually going well and that two sides might be able to reach a deal. And then the leaking of that memo ended everything. Emil, Michael put out a tweet where he said, like, negotiations are over. This is over. And so Dario issued a public apology in a blog post where he wrote, quote, it was a difficult day for the company. And I apologize for the tone of the post. It does not reflect my careful or considered views. It was also written six days ago and is an out of date assessment of the current situation. Now, Matt, I know that I was mean to you last week, but it was six days ago. I'm a new man. Don't even worry.
- [11:26]
A
Greg, I know it was a totally different you. Everything you said, totally different. You don't believe today.
- [11:31]
B
Yeah, you know, I've grown. I've changed.
- [11:34]
A
So much has changed.
- [11:35]
B
Anyway, so that's, that's kind of like the, the key play by play action of where we are until those lawsuits which were filed on March 9th.
- [11:45]
A
Yeah.
- [11:45]
B
Right.
- [11:45]
A
So you said there are two different lawsuits that were filed. Can you talk a bit about why there were two? And then we can sort of get into the weeds of each lawsuit.
- [11:53]
B
Yeah. So the existence of two lawsuits directly pertains to Anthropic's legal strategy in fighting this determin by the Department of War and the broader action by President Trump and Pete Hegseth on social media. So in the actual letter from Hicks Hegseth to Amadei, it says that he is invoking the authorities under 41 United States Code 4713. Right, section 4713. Now what's interesting is that section 4713 derives from federal law. That actually is a case in which Congress has limited the opportunities for judicial review. Now, this is not as crazy as it sounds. Right. Let's, let's imagine the most extreme straw man case here for a second. Let's say we're not talking about Anthropic. Let's say we're talking about Huawei. And let's say that the way that the Department of War knows that Huawei is a supply chain risk is that they have, you know, bounced a satellite laser off the moon to hack into Xi Jinping's phone. And they tapped a call between Xi Jinping and the chairman of Huawei, Ren Zhengfei, in which Ren Zhengfei says, bwahaha, I have infrastructure inside the DoD and I'm going to use it to destroy them. Right. Well, you definitely don't want to be in a situation where the Department of War has to in public court filings say we know how to hack Xi Jinping's phone by using this awesome satellite which bounces a laser off the moon. Right. They don't want to disclose the existence of that technology. They don't want to disclose the existence of that surveillance operation. They don't want to disclose what they've learned from that surveillance operation. So those are all pretty reasonable justifications in which Congress has written the supply chain risk authority such that the judicial review is limited to only one court, which is the United States Court of Appeals for the District of Columbia Circuit. That's a court that actually is equipped to handle classified information. And in a way that I would be kind of scared if I was Anthropic, it's a one way presentation of evidence so the Department of War can present classified evidence to the court. Anthropic has no right whatsoever to review that classified evidence. So again, you know, using the strawman Huawei example, why might that be a good idea? Well, it's not enough that the evidence merely not be made public. It has to be not made available to Huawei. Because if we know that Huawei is a front for the Chinese intelligence services and military, you don't want to disclose to them, even if you don't disclose publicly. In fact, disclosing them is just as bad as disclosing publicly because you're directly giving it to the adversary. So that, that the, the, the, the law in question restricts Anthropic's opportunities for judicial review. For this authority that was invoked in this letter under section 414713 and basically says it has to go to that court. And the first thing that that court has to decide is whether or not the Department of War has a reasonable case here for restricting judicial review. Now, the second court filing, which is the Northern District of California court filing, I mean, for those of you watching on video, I'm holding them both up here. This one is 48 pages, the other one is 13 pages. And that's because the meat of Anthropic's complaint, the much more diverse set of complaints and requests for legal redress are in this Northern California court. And here's the interesting legal argument that Anthropic is basically making. They're saying, okay, the law says that harms related to section 4713 have to go through the United States Court of Appeals for the D.C. circuit. Fine. But we have been harmed in Many other ways besides just that. 4713 filing that and the harms come from the truth social post, from Secretary Hegseth's tweet, and not just from the official letter. Now, what's interesting here is the evidence in this is like extraordinary. It's a. Like what really came to mind as I was reading it, and again with the caveat that I'm not a lawyer, what came to mind is that scene from Wolf of Wall street where the FBI is going up to Leonardo DiCaprio and they're basically saying this is a Grenada. You know, it's like the US Invasion of Grenada where our military is colossal and you're an insect and the war is basically over the millisecond we decide to fight it because we're so much more powerful. Like there is a lot of really powerful argumentation and evidence in this Northern district of California. But the first argument, the first argument that Anthropic has to win is that they have the right to even have a case in Northern California.
- [17:47]
A
What are they asking for in Northern California? Because the first is a judicial review case. What is this one?
- [17:53]
B
Yeah. So again and again, what they're asking for is what's called a preliminary injunction. And a preliminary injunction is where you basically argue before the court that the harm that you are suffering is so severe, so urgent and so unredressable that the action in question has to be stopped even before the trial is over. So what they're sent, they're essentially saying is like, look, we're probably going to be at war in court. I shouldn't say at war given the context. We're going to be fighting this out in court for years. The those years could be life or death for Anthropic. And their arguments are so weak and our arguments are so strong that you should not let the executive branch succeed in taking these actions even for the duration of the court proceedings. Now, maybe at the end of the, you know, several years, there will be some kind of finding that the Department of War is correct, that Anthropic is wrong, and the preliminary injunction will be overturned. But what Anthropic is basically saying is that the evidence in this case is so one sided that you should block the executive branch's actions even before the overall court proceedings have concluded. And that's why you see in this filing a constant reference not just to dramatic harms, but to irreparable harms. So what that means is even if four years from now, Anthropic wins the court case and The Department of War has to pay Anthropic, let's just say, the total value of the contract for all of those years that Anthropic was unable to perform any work and receive any revenue. Let's just say the Department of War has to give it to him then. Well, what Atropic is basically saying is even that would come nowhere close to redressing the harm that we will have received for this period of time. Which is why we need a preliminary injunction. Yeah. And that fight, that fight over, you know, is this court case going to be heard and then also is a preliminary injunction going to take place? That could happen in a matter of weeks. Right. Like that could be somewhere in the two to eight weeks timeframe. And if that occurs, it will occur on what is kind of favorable home turf for Anthropic. The Northern District of California is chosen because that's where Anthropic is located. Those are justices that historically have been more skeptical of this administration and the argument that it, the arguments that it tends to make in court. And here's the other thing. Right. What if, what if the United States District Court of California finds that they do have standing to hear this case and that the suit can proceed and there's a preliminary injunction, while simultaneously the United States Court of Appeals for the District of Columbia Circuit fines. Actually, we have exclusive authority in this matter because the only real things that have taken place are these 4713 actions, and we have exclusive jurisdiction for that stuff. So now you'd have two conflicting court rulings.
- [21:36]
A
Yeah.
- [21:37]
B
And this very plausibly could be headed to the, to the Supreme Court. It's really. Yeah. I mean, you've heard me going on for maybe too long about it, but like, it expresses the stakes here and the complexity here and just how much hinges upon these sort of legal determinations.
- [21:58]
A
Yeah. Well, I want you to go on a little bit longer and tell me about some of the claims that Anthropic makes in the second lawsuit, the one that was filed in California.
- [22:08]
B
Yeah. So I think what's, what's really interesting is, you know, you've heard a lot of people in the media being quoted anonymously saying stuff about, you know, anthropic is used for X in the military, anthropic is used for Y in the military. Or, you know, Pete Hegseth said this in a private meeting. What's interesting is that, like, all of this is submitted under oath. I think. I think it's submitted under oath. It's definitely illegal to lie knowingly lie in a court filing like this. So you get attested quotes of what Pete Hegseth said in that private meeting with Dario Amadei. Now, it's not impossible that Anthropic could lie in this court filing about what Pete Hegseth did, but they're. They're lying in a way that could put people in jail or certainly to be fined or found in contempt of court or. I don't know. But, like, the point is, it's a much bigger deal.
- [23:06]
A
Incentives to lie.
- [23:07]
B
Yeah, it's a much bigger deal to lie in a court filing than to lie in a tweet. And so what did, according to this court filing, Pete Hegseth say? He said that Anthropic's military capabilities are exquisite. That's a quote. Exquisite. And that the Department of Defense, this is not a quote, loves what they're getting from Anthropic. That it is like the leading overall AI capabilities. And then Anthropic has been really important, you know, in a ton of very important military operations. So that's the kind of thing that, like, Dario was not going to go on cable news and say, hey, Pete Hegseth, you know, said that we were exquisite. But now that it has gone to court, all of those kinds of details are relevant, factual information for who should win this court case. Another thing that I think is so interesting and powerful in this is, again, a lot of this court filing for the California case is fighting for standing, fighting for the right to have this heard. So they're pointing out, like, damages that they have received just from the Truth Social post or just from Hegseth's tweet. And I think those are genuine. I mean, I've heard from people who were traveling abroad and meeting with foreign government officials, and those foreign government officials saying, oh, I guess we can't work with Anthropic anymore now that they're subject to American sanctions because their. Their mental model is like Huawei or buying Russian oil style sanctions. And so the key here is that Anthropic can be harmed by tweets, even if those tweets exceed legal authorities, because the perception of those legal authorities can themselves perpetrate harm to the company, and those harms might be irreparable. The other thing that Anthropic is doing is they're making big claims about free speech, which folks will remember in US Law, it is totally routine to refer to companies as a person and that such persons can have free speech rights. And Anthropic is basically saying, you know, Their defense of AI safety, their defense of AI security, their advocacy for or against various laws. Even if the administration doesn't like them, they can't punish them just for expressing those views. And they cite court cases where even the temporary chilling or punishment of free speech can be an irreparable harm. Which again, as I said before, when you claim that you're an irreparable harm, you're making an argument for a preliminary injunction.
- [25:43]
A
Yeah, and I think there's also been some interesting. There's some interesting stuff in there about how this has hurt Anthropic financially already, as well as the potential for it to hurt Anthropic severely. I think actually the Anthropic cfo, Krishna Rao, submitted a filing talking more about this. I was hoping you could expand a little bit upon that as well.
- [26:05]
B
Yeah, it was really interesting. They said that Anthropic, since the founding of the company, has received $5 billion of revenue, has spent $10 billion acquiring compute capabilities to train its models. And I struggle to understand how those numbers could be true. Other reporting about Anthropic has talked about how just in the past year their revenue has grown by more than Palantir's annual revenue, for example. And Dario has talked repeatedly in interviews about Anthropic's revenue growing 10x every year for a while. So I don't believe that Anthropic is lying in that. I think what's happening here, and this is again me speculating, is a difference between revenue as it is recognized on formal accounting statements. So just for folks who never went to business school, you can't recognize revenue until you finished delivering the service. So if you, if you sign a contract with a company for 10 years and it's a $10 billion, even if they're paying you $1 billion a year for 10 years, depending on the nature of the contract, you might not be able to recognize the delivery of all of that revenue until the completion of the 10th year. Definitely there's circumstances in which you can recognize revenue on an ongoing basis. But I'm just getting at that general principle is probably why the stated revenue number is lower than what has been reported elsewhere in the media and talked about by Anthropic. It's a difference between revenue that is recognized formally versus cash in hand versus accounts receivable, blah, blah, blah, blah, blah. But Anthropic is basically saying that the harm is not just the harm that we have received from a loss of revenue from the Department of War on this contract that has a $200 million ceiling. Nor is it just a harm from all the additional revenue that we might receive from the Department of War. Nor is it the harm that we've received from, like all federal customers. They are citing the harms, as Pete Hegseth said them in the tweet, which is to say that any commercial entity that does any business with the Department of War cannot do business with Anthropic. Now, what's interesting is, is that that's actually farther than some key companies are going in their own legal reading of the circumstances. So Google, Microsoft and aws, Amazon have all said that their interpretation of the events is that this memo, the memo that the letter that Hegseth sent to Anthropic, that that is like the legally binding document, and that document is more narrow. It just says that it covers all Department of War procurements. So basically, when Microsoft is working on something that's going to be shipped to the Department of War, they can't use Anthropic. When Microsoft is working on something that's going to be shipped to, you know, U.S. steel Corporation, they can use Anthropic. That's like what Microsoft's lawyers are interpreting.
- [29:50]
A
Yes. So unlike, unlike Pete Hegseth's tweet, they're saying, look, we can use Anthropics models in commercial applications. Whereas, whereas Pete Hexseth had explicitly said that they wouldn't be able to use it. Okay.
- [30:02]
B
But what Anthropic is claiming is that not all of their customers are acting under those assumptions. They have customers. And even Lockheed has said this. Lockheed Martin, which is a major US Government contractor, has said that they're pulling Anthropic out of their technology stack everywhere. Now, like 99% of Lockheed Martin's business is with the U.S. federal government. But the point remains is that that is evidence of some kind of chilling impact coming from the President's Truth social post and from Hegseth's tweet that goes farther than that formal letter. And again, because the letter is the 1 invoking the 4713 authorities, the 4713 authorities have the opportunity to limit judicial review to the D.C. court of Appeals. It is very important for Anthropic that they show harm that goes from the government's actions that went beyond those 4,713 actions, because that is what gives them permission to sue in a different court on more broader grounds with more opportunity for judicial review.
- [31:15]
A
Yeah, and there's this also one other part to this. These lawsuits that I want to talk about which is that Anthropic is not only arguing that the supply chain risk designation was illegal. It's also arguing against Trump's directive and against potential executive order that all government agencies.
- [31:32]
B
They cite here, that hhs, the Department of State, has. Has ordered agencies to rip Anthropic out.
- [31:37]
A
They've already started. Yeah. Moving Claude out of the ecosystem.
- [31:40]
B
And supposedly, according to Axios, an executive order is coming on these grounds. Now, if that executive order comes out again, that's probably not going to be under 4,713 authorities, which would give, you know, Anthropic another opportunity to sue. So this is why I say that if Anthropic can get standing, my amateurish read of their filing is that they have a strong case and that they are likely to win if they can win the right to be heard in court.
- [32:09]
A
Yeah. Well, Greg, you and I are think tank bums, and we spend all day scrolling on Twitter and substack, and you send me your Twitter posts in the early hours of the morning, and I read them because I want to know. But for those in our audience who, unlike us, don't have that time to just be scrolling on these platforms where so much discourse is happening, can you break down what people are saying? Because there's been so much online from various experts in this space who have a lot to say about this.
- [32:37]
B
People who are watching on YouTube will notice that I'm, like, scrambling this jumble of papers that I have in front of me, because really, there is just so much information coming from so many different angles across all of this. And so our task as a humble podcast host is to take all of that complexity and to digest it. There's a lot of that has been said that is worth quoting here, but I'm going to filter down to a few that I thought was interesting. Number one, I would say comes from Senator Ted Cruz, who, on cnbc, I think it was today, and we're recording this on March 10, he said, quote, I'll confess I have not seen a basis laid out for why the government would be prohibited from using Anthropic. Claude is one of the many AI tools that can be very helpful. I don't think government should be picking winners and losers. So that's interesting. I mean, you've got now Republicans in the Senate not viciously criticizing what the Trump administration is doing here, but basically saying that they don't see it. And again, that's actually important in and of itself because part of the anthropic legal filing relates to violations of the Administrative Procedure act, which basically says. Says that executive actions done in violation or sorry, in the absence of legislatively required procedural steps are invalid. And one of the required legislative steps was notifying Congress before you take the relevant action. So, like, if Ted Cruz is like, I haven't seen credible justification. Well, if that's because the Department of War never attempted to give them one, that would be, you know, something. They could really do it. Sure. So that's. That. That's one thing that I thought was interesting because there's already been quite a, quite a lot of vicious criticisms from Democrats in Congress. But the fact that you're seeing it come from close Trump allies like Ted Cruz.
- [34:43]
A
Yeah, Ted Cruz was, he was very involved on the anti state regulation agenda.
- [34:49]
B
Right.
- [34:49]
A
Earlier.
- [34:49]
B
Yes.
- [34:50]
A
Earlier in 2025, which anthropic was an opponent for. And so you would have thought that they would be on opposite sides of things. So it's pretty significant he's coming out here and saying this.
- [34:59]
B
Yes. And Ted Cruz also notably has fought with this administration on free speech issues previously. So, you know, Anthropic is suing that their speech rights have been violated. Ted Cruz went on a lot of conservative podcasts, including I think, his own podcast, because he has one, and said that when Jimmy Kimmel, a late night talk show host on abc, was temporarily suspended for comments made that the Trump administration viciously criticized and that the fcc, you know, threatened some retaliation for. Ted Cruz was one of the big defenders, not of what Jimmy Kimmel said, but of his right to say it. And that was, I think, an important moment in terms of, like congressional Republicans pushing back on the administration. So here is Ted Cruz again being a notable voice on this topic, pushing back against the administration. A second voice that I think is of course important is Emil Michael, the undersecretary of war for research and engineering. He went on the all in podcast, which is an influential podcast in the tech right, and he had a question where he was asked in response to, like, why Anthropic and the Pentagon didn't simply end the contract. Emil said, quote, I don't view it as punitive. And I'll tell you why. If their model has this policy bias, let's call it based on their Constitution, their culture, their people, and so on. I don't want Lockheed Martin using their model to design weapons for me. If Boeing wants to use it to build fighter jets, I can't have that because I don't trust what the outputs may be because they're so wedded to their own Policy preferences. Very, very interesting there because like that, it's really not the same justification that we heard from the administration elsewhere. And then finally, I think one thing that is super interesting, this is something that Michael said. Emil. Michael also did an interview with Pirate Wires, which is a new technology journalism outlet that has a lot of right leaning political inspiration. And he said this, he was asked, you know, by the reporter, has the bridge between Anthropic and the Pentagon been totally burned? Here's what he said. Quote, I have a responsibility to the Department of War and if there was a way to ensure that we had the best technology, I have no ego about it. But they have to re earn the trust, be reliable, prove to us they could be a long term partner. Because as we use AI more and more in what we do, like every big organization does, we've got to have long term trust. That's like the basic principle. And they'd have to prove that. But if it was provable, if that was provable, I can't speak for the Secretary or the President, but I would be open minded. I mean, look, I'm a deal guy,
- [37:47]
A
he's a deal guy.
- [37:48]
B
So here's the thing. His open mindedness and Dean Ball, who's a former senior advisor for the Trump White House, the second Trump administration on AI stuff, he's pointed out that like this claim that they're still open to a deal deeply undermines the security justification. Right? Like go back to the Huawei strawman example I gave, right, where there's super secret classified evidence proving that this company is super duper unreliable. Well, if there was such super duper classified evidence proving that Anthropic was unreliable, why would Emile Michael still be open to bringing them back in? And Anthropic makes another key point in this filing, which is that it's not just that Anthropic is still being allowed to provide critical services to the Department of War and the US Military for an ongoing conflict in Iran where lives are on the line, it's that not a single security clearance has been pulled for Anthropic. Right? So imagine, let's go back to our Strawman example, right? Huawei is in the Department of Defense's infrastructure. And the Department of Defense says, oh wow. Based on this super secret classified evidence, we've determined that Huawei is a security risk. But we're going to let every Huawei employee keep their security clearance for six months and we're going to, we're going to. Well, the security clearances might even go longer than Six months, who knows? But, like, we're gonna let them continue doing work in the middle of a war for six months? I mean, it's a very, very tough case to make, I would say. I would not want to be the lawyer defending that position in court. And that's kind of where the Trump administration finds itself.
- [39:38]
A
Yeah.
- [39:38]
B
So let me say also what Dean Ball said, because he argued pretty dramatically on his own substack in the Atlantic and on the Ezra Klein podcast. It was really dramatic. Here's what he said in substack. Quote. I consider the events of the last week a kind of death rattle of the old American republic, the outward expression of a body that has thrown in the towel. And it goes on to say, essentially the United States Secretary of War announced his intention to commit corporate murder. The fact that his shot is unlikely to be lethal, only very bloody, does not change the message sent to every investor and corporation in America. Do business on our terms or we will end your business. Wow. That's quite a quote from a man who, this time last year, worked at the White House.
- [40:33]
A
Yeah. And not only worked at the White House, but had a huge role, one
- [40:37]
B
of the principal authors of the AI Action plan. AI Action plan, which is still the flagship Trump policy on AI.
- [40:44]
A
Yeah. And I do want to note that, like, for most of this conversation, we've been talking about one question, which is, is the supply chain risk designation? Does that make sense? Is that legally viable? There's still these other questions that we talked a lot about in our. In our February 25 episode about whether companies should be able to restrict the government from using their systems, just in general, whether whether it's wise for the.
- [41:05]
B
The government to go. Go after companies in this way. And I would refer folks to also encourage them, if they can tolerate my voice longer, to listen to the interview I did with Ben Thompson on Checkery, where we went into this at great length. But I would basically just say, I mean, I was in the government in the aftermath of the Project Maven fiasco in 2018 when Google pulled out, and it really looked like there was going to be this dangerous divorce between Silicon Valley and the military over the use of AI in particular. And I remember how hard it was to try and convince leading AI scientists or leading companies to work with the Department of Defense. And one of the things that I would always try and say is like, hey, come on in, the water is fine. Your company can have a good business here. Your company can do work they're proud of here. Your company can work here in A way that is consistent with your ethical values, with your company's principles, and, of course, American law. And now, if I was trying to make that pitch again, I really feel like the message that the Department of War has sent is that anytime any company dips their toes in the defense contracting waters, we reserve the right to grab your ankle and pull you all the way in or even to just drown you if we don't like you. And that's a terrible message, you know, when you really do need the private sector. If we're to win this AI race against China, if we're to do all the important things that we need AI to do for our security and prosperity, this is a terrible precedent to ascend. And I say that as somebody who is deeply sympathetic to many of the arguments the Department of War is making about how they must be in charge of when technologies get deployed, how they must have operational control over their technologies. I say that even being sympathetic to many of the Department of War's arguments.
- [43:12]
A
Yeah, yeah. Well, I think that's a good place to wrap up on this Anthropic Pentagon stuff. I mean, more is to come for sure, but we'll cover that when it happens. But I want to move on to another topic that you've alluded to several times previously in this recording, which is how AI is being used in the Iran war. I mean, specifically, we know that Quad is being used in the Iran war. So what do we know about how it's being used and what the impact has been?
- [43:35]
B
Yes. Okay, so if you thought I was doing a lot of quote reading before, brace yourselves, because here we go. As I said, this filing that Anthropic has done, especially the Northern California one, it is a banger, and I encourage everyone to read it themselves. But we've got a really interesting pattern of facts. Some of this was already available in public, and Anthropic has just helpfully sort of synthesized all of it. Other stuff was not public or was only, you know, provided by anonymous sources on background and now is public. So there's a few things. Here we go, from page 13. Anthropic has also developed specialized Claude gov models tailored specifically for the national security context. These models have been built based on direct feedback from national security agencies to address real world requirements like improved handling of classified information, enhanced proficiency in critical languages, and sophisticated analysis of cybersecurity data. Claudegov models undergo rigorous safety testing consistent with Anthropic's commitment to responsible AI. Then it goes on, and this is, I think, really juicy to make CLAUDE more useful for the military and intelligence components of the federal government. Anthropic does not impose the same restrictions on the military's use of CLAUDE as it does on civilian customers. CLAUDE government is less prone to refuse requests that would be prohibited in the civilian context, such as using CLAUDE for handling classified documents, military operations, or threat analysis. Anthropic's terms in its existing contracts with the government also recognize the government's unique needs and capabilities. For example, Anthropic's government specific addendum to the usage policy permits CLAUDE to be used to analyze lawfully collected foreign intelligence information, which would not be permitted under the usage policy for civilian users. So, I mean, you've got a lot there, right? It basically says out loud what a lot of people have been saying in the background, which is the intelligence community is hoovering up a ton of information and they're feeding it into claude, and what they're getting on the back end is awesome. So let me just like explain what this means in walking up the ladder of using AI capabilities as AI has become more capable. So when I was in the military, AI was a synonym for machine learning, and nobody was using a large language model anywhere to do anything interesting. So what were machine learning capabilities that might be interesting? Well, we've talked about this before on the podcast. You know, in the context of Ukraine, there's a lot of intercepted Russian radio communications and what machine learning with speech recognition, which back in 2017, nobody questioned that speech recognition was AI, nobody questioned that facial recognition was AI. Now these, these terms have sort been demoted and they're no longer AI, but at the time they were seen as really cool, cutting edge technologies. So speech recognition allows you to take all that intercepted Russian audio, which normally you would have to have a Russian linguist listen to it at the rate of 1 minute per minute to convert that spoken stuff into text stuff. And the text stuff is a lot more useful because you can then search it right, for critical keywords, et cetera, et cetera. But what this, I think this is hinting at is that we're no longer just using the AI to convert audio into text. We are now using the AI to like read all of the text at once, or read all of the audio at once and connect dots in really important ways. So we can sort of say like, hey, these conversations, you know, appear to be talking about some important location of something that we might want to blow up. These, you know, intercepted odd audio talks about patterns of life that we might be able to use to infer Importance of things. Again, I'm speculating a lot here, but it appears to me to suggest that they really have plugged the AI into some really powerful intelligence capabilities in a way that matters. And here's what they say. They go on to say they have allowed for the use of CLAUDE to support government operations such as the rapid processing of complex data, identifying trends, trends, streamlining document review, and helping government officials make more informed decisions in time sensitive situations. I realize some of that sounds pretty blase, but I can kind of read between the lines here like they're doing important military work, they're doing important intelligence.
- [48:35]
A
We also have some good reporting, I think, from the Washington Post that goes a little bit more, more in depth about, like how this has helped the military operation in Iran so far.
- [48:45]
B
Yes. So one thing that I think is important to know, note just about the overall situation in Iran is that Iran has a lot more missiles and what you might call pseudo missile drones than America and our allies have missile interceptors. So what that means is if Iran is permitted to shoot all of their missiles, that's a really bad day for America and its allies because we are nowhere near being able to shoot down everything that they can shoot at us. So just from a military strategy perspective, it is very, very important for the United States military to blow those missiles up, to blow those drones up on the ground, not in the air. And similarly, it is very, very important that we blow up the launchers for those missiles and the, for those drones on the ground before they can start shooting them in the air. And that is, that is like the sort of critical phase of the war that we are in is blowing up all the launchers we can find and blowing up all the missiles that we can find. The problem is that Iran has done a lot of stuff to hide where their missiles are stored, to hide where their launchers are. And a lot of their launchers are mobile. Right. They're like a truck. So they drive out of a cave or they drive out of some kind of underground bunker and then they're moving very, very fast. And the United States military for a very long time has had an intelligence problem of an asymmetry between the amount of data that we can collect and the amount of data that we can analyze. And when I was in the military, you know, AI for analyzing satellite images or AI for analyzing drone images, it was used for comparatively humble use cases. Right. We would say there are a hundred vehicles in this picture. Or, you know, we would say this picture is of empty ocean. This picture has A warship in it. Human analyst, I recommend you look at these pictures first. I think what this is suggesting in terms of like the critical capabilities that Anthropic is delivering is like you have these big drone platforms, you have these big satellite reconnaissance platforms. They're collecting a incredible amount of imagery, data and a lot of other kinds of intelligence, such that even if we don't know where a bunker is, if a transporter electric erector launcher, basically a truck that can launch missiles, pops out of the ground, anywhere that a drone is looking at, the AI catches it immediately in a way that would be very, very difficult for the, for even trained human eyes to notice, just because there is this ocean of full motion video data that is being captured and a very finite number of trained eyeballs to look at all of that data. And that is so powerful because those transporter electro erector launchers, we don't know where they are, but as soon as they start moving, you know, it is a time critical decision to blow them up before they start launching their missiles. And so basically having analytical capacity that matches our data gathering capacity, that has been a crisis in US Intelligence for over a decade. And I think what they're saying is that large language models, especially multimodal models, hooked up to Palantir, hooked up to the latest and greatest version of Project Maven enabled by Claude. We're sort of finally getting to that fantasy of being able to analyze everything that we can look at in real time. And that is a sea change for the Department of War. And it explains why some of the anonymous quotes that you've seen from government officials of, you know, Anthropic. The problem for Anthropic is, quote, they really are just that good. You can kind of get a sense of why these capabilities are so desperately important. And it comes back to some of the quotes that this, this court filing talks about, like what the Department thinks about Anthropic. So here I'm going to read again. Quote. Even during the recent negotiations, the government has repeatedly and publicly praised Claude's capabilities. Chief Technology Officer and Undersecretary of War Emile Michael, while describing the dispute with Anthropic, Anthropic explicitly characterized Anthropic as one of America's, quote, national champions in AI. In the February 24 meeting with Dr. Amadei, Secretary Hegseth described Anthropic's technology as having, quote, exquisite capabilities, end quote, and stated that the Department would, quote, love to work with Anthropic. And here's another quote, but this one, they're actually just reporting what was already said in the media. Department officials have even expressed concerns about the consequences of losing access to Claude. Describing the dispute between Anthropic and the department, one official stated that, quote, the only reason we are still talking to these people is we need them and we need them now. The problem for these guys is they are that good. End quote. Wow. And that you come back to like, why did they have this six month grace period? Why did they have this six month off boarding period? Because they really need these capabilities because they really are powerful and delivering amazing advantage for US national security. We've talked about the imagery possibilities of a technology like this. We've talked about the SIGINT possibilities of a technology like this. But one thing that I think is also important is the connecting of all the dots. Right? It's like having an AI model that can speak the language of SIGINT and speak the language of GEOINT and speak the language of analyst and, and assembling that information in meaningful time frames and having an audit trail for how, you know, so you can check its homework. That is really just shocking. And it's the kind of thing where, gosh, we were, we were, we wanted that future so badly when I was in the Department of Defense and we, we didn't have it, we didn't have anything close to it. It's just, it's just really shocking to read about it in this court filing of how far we've come.
- [55:24]
A
Yeah, well, I think maybe that does it for the anthropic court filing readings. But I did, I asked about the Washington Post and I got more anthropic court filings, which is forgivable because of how interesting all the anthropic court filing stuff is. But I want to get to this Washington Post article because I think there's some really interesting stuff in there about specifically Maven Smart Systems. But can you talk a little bit more about what they're talking about in this article?
- [55:49]
B
Yeah, yeah, yeah, you're totally right to bring us back here. The filings are really juicy, but there's also some juicy stuff coming out in the journalism right now. So let's just use a few key quotes from some important Washington Post articles. So, quote, in order to strike a blistering 1,000 targets in the first 24 hours of its attack on Iran, the US military leveraged the most advanced artificial intelligence it's ever used in warfare. A tool that could be difficult for the Pentagon to give up even as it severs ties with the company that created it. And then it goes on to say, quote, the military's Maven Smart system, which is built by data mining company Palantir, is generating insights from an astonishing amount of classified data from satellites, surveillance and other intelligence, helping provide real time targeting and target prioritization to military operations in Iran, according to three people familiar with the system. And I'm going to keep going because I think it's just that juicy. Quote. As planning for a potential strike in Iran was underway, Maven powered by Claude, suggested hundreds of targets, issued precise location coordinates and prioritized those targets according to importance, said two of the people. The pairing of Maven and Claude has created a tool that is speeding the pace of the campaign, reducing Iran's ability to counter strike and turning weeks long battle planning into real time operations.
- [57:14]
A
So this is kind of confirming what you just talked about, right? Like that's exactly what seems to be happening.
- [57:19]
B
Yeah. So you know, the court filing is talking about the types of activities, how and putting it all on the record and the sort of legal equivalent of under oath. But you've also got this anonymous sourcing which is putting some flesh on the skeleton. Right, that we've gotten there. And it's not just, you know, my speculation. Right. This is anonymous sources in the Pentagon, but 1,000 targets in the first 24 hours, prioritizing those targets, giving locations, and if you know anything about Palantir, one of the things that people loved about Palantir's work and has been true for a while is the extent to which it gives an audit trail. So like based on these sources, we are making this judgment. So it's not just like you're getting this, you know, input from the LLM and it says, trust me, you need to blow this up. It's like, here's what you need to blow up and here's why I think that. And if you want to check my homework and tell me I'm wrong, you know, here's, here's how you would go about doing that. Yeah, it's a really powerful, exciting capability. You can get why NATO allies are clamoring to get access to this technology. You can get why Japan is clamoring to get access to this technology. You can see why the United States and China and everyone else, and I've been saying this for a decade, view AI as foundational to the future of military power. But I need to amend that statement. It's no longer true. True, AI is foundational to the present of military power. And if you're going to fight a War against a country that has this capability and you don't. That's a bad day. That's a scary bad day.
- [59:00]
A
Yeah. And there's other ways in which we've seen the new focus on AI emerge through this, this war in Iran. One of those is that Amazon data centers were targets of drone strikes, right?
- [59:12]
B
Yes. And I don't know that we can really.
- [59:15]
A
Yeah, the connection.
- [59:16]
B
I don't know what we can make of this yet. It's a really interesting fact. We don't know what was in Iran's head as they did this. Like, okay, so here's the reporting that we have, which is from a March 2 Reuters article. Some of Amazon's, quote, data centers in the United Arab Emirates and Bahrain were damaged by drone strikes in the Middle east conflict, disrupting cloud services and making a recovery prolonged. So what does that mean? Does it mean that Iran recognizes data centers as critical infrastructure or incredibly valuable assets worth blowing up? Does it mean that Iran has an opinion on whether or not those were classified AWS capabilities? Because AWS makes a classified capability available to both the Department of War and to the intelligence community. We have no sense of what flavor of AWS capability were targeted in these strikes and even if they were targeted, whether that was successful. But I do think it points out that these are strategically essential assets. I mean, there's a reason why companies and countries are willing to drop $25 billion on a data center or $800 billion on capex data center Capex for a given year. And maybe the Iranian specific attack in this instance was not an especially thing. But even if that is the case, I think in future wars it will be totally routine to try and bomb data centers.
- [60:48]
A
Right, right.
- [60:48]
B
Because of the AI capabilities they provide.
- [60:50]
A
Yeah. And I think one of our colleagues here at csis, we have to give a shout out to wrote an article or a commentary about this before the Iranian war started, before these data center strikes happened. So I just wanted to give a shout out to Mona Yakubian here in her commentary. I don't know if you want to speak any more about.
- [61:06]
B
No, no, no, you take it away. Yeah, so I haven't read this.
- [61:08]
A
Yeah, well, she wrote, quote, with conflict looming in the Gulf, the region's nascent AI infrastructure could be imperiled, further raising the stakes of regional war. So she predicted that the AI infrastructure could be imperiled. And then she made a comparison that I think is interesting to oil. So she wrote, in previous conflicts, regional adversaries such as Iran and its proxies targeted pipelines, refineries, and oil fields in the Gulf partner states. In the compute era, these actors could also target data centers, energy infrastructure supporting compute and fiber choke points. So again, it's notable that she wrote this right before these strikes happened. It's not 100% clear that these strikes are evidence of that, but they surely could be.
- [61:50]
B
Yeah. At a minimum, it's a plausible interpretation of the fact pattern and you could be forgiven if you said it was the best interpretation and the likely interpretation of the fact pattern. So yeah, kudos to her. And and yeah, it really brings me back to if we're building these ultra powerful capabilities, we really do have to think about defending them. It's a part of national security. And in the same way that we have to think about protecting US Military basing infrastructure abroad if companies are going to put critical parts of their ability to generate economic revenue or even support war fighting capabilities. That's all stuff we have to think about defending now.
- [62:31]
A
Absolutely. Yeah. Well, I think this is a good place to wrap it up, Greg. Thanks again for your expert analysis and thanks to our audience for tuning in.
- [62:39]
B
Yeah, and thanks everybody out there who made it this far. I know this was a doozy of a podcast, but just so much juicy information coming out I thought we had to break it down.
- [62:49]
A
Absolutely. Thanks.
- [62:53]
B
Thanks for listening to this episode of the AI Policy Podcast. If you like what you heard, there's an easy way for you to help us. Please give us a five star review on your favorite podcast platform and subscribe and tell your friends. It really helps when you spread the word. This podcast was produced by Sarah Baker, Sadie McCullough and Matt Mand. See you next time.