Pentagon CTO Emil Michael: Anthropic is still blacklisted, but Mythos is a separate issue

A landmark agreement has several major AI leaders agreeing to deploy their AI capability on the Pentagon's most secure and classified networks. This news is just out today. Those companies include SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, and Amazon Web Services. Joining us right now to talk more about it is Emil Michael. He is the Pentagon's Chief Technology Officer. And Emil, thank you for being with us today. Let's describe what went into this. And why you are now going to be deploying so many different AI services and capabilities throughout the department?

Well, what we've learned since we started this effort at the Department of War is that it's irresponsible to be reliant on any one partner. And when we learned that that one partner didn't really want to work with us in the way we wanted to work with them, we went out and made sure that we had multiple different providers, both open source which is a new, a new effort here at the department, and the proprietary model companies and the infrastructure companies like Microsoft and AWS, and got them to agree to sign up to work with us on classified networks to make sure we had diversity of supply. We had the best of the best, and we had multiple different paths with open source and proprietary.

Look, I understand wanting to diversify and not be reliant on one supplier for anything. That makes sense. When you're using this many, do you risk having chaos from trying to figure out how to implement all these different things? How do you manage that? And then you've got employees at Google who have asked their company to reject this offer or this, this deal with, with your department. What, what do you think about that? Because I know you're referring to Anthropic and the issues that you had there. They're, by the way, not on this list of companies that you're going to be working with on this.

Yeah, that's right. They're not on the list. Basically, no, we don't have confusion because what we do is we try to take each model for its strengths and use it in those particular use cases. And you think each of these models is trained on somewhat different data. If you look at Google, you know, they have email, YouTube, Nest cams. They have a different sort of, you know, capabilities. They may all converge over time. But we're going to use them in the right ways. And we've already deployed SpaceX's XAI and OpenAI, and we're benchmarking that against what we've had previously. And then open source is a brand new vector for us, which is very important because Chinese open source models have infiltrated a lot of different companies. We want to have an American alternative to that. As it relates to the Google message, I think it was less than 0.1% of their employees wrote a letter. But thanks to the leadership there that learned from the lessons of 2018 with Maven, that serving the US government and being a partner of the US government is really important. And they were, they were proud about that. And I'm proud that they stood up and agreed to work with us in the way that made sense for us.

Emil, it's Andrew here. A couple of quick questions. What is the state of play with the conversations taking place around Anthropic with the Department of Defense? Of course, this also comes as the NSA appears to be working with them. And I know there's been lots of back and forth about other parts of the government working or not ultimately working with them.

Yeah, with Anthropic, their supply chain risk, we've given the different department components 6 months to transition off. If there's some exceptions, you know, we'll make them because the complexity of integrating into some of these networks is really high. That being said, I think the mythos issue that's being dealt with government-wide, not just at the Department of War, is a separate national security moment where we have to make sure that our networks are hardened up because that model has capabilities that are particular to cyber, finding cyber vulnerabilities and patching them. And from a national security standpoint, you always have to look at those things. And NSA and Commerce evaluates all frontier models, including Chinese frontier models, to see what the capabilities are at the edge.

When we last spoke, you had made a comment about Anthropic having a constitution which you thought was at odds ultimately with what the government was hoping to do and the freedom that the government should have. And one of the things that I wanted to ask you about is OpenAI. Well, I don't think they call it a constitution, but or Gemini, they might not call it a constitution per se, but in a very similar way, they have built effectively instructions and guardrails into their systems to do certain things and not do other things. And some of those things are almost identical to the things that Anthropic has built into its guardrails, if you will. And so I'm hoping maybe you could decipher or explain what the distinction is in your mind.

Yeah, you're— I think you're mixing up guardrails and, and, you know, how the model is designed around core principles. Guardrails. We want guardrails at Department of War in some ways, right? We don't want some of the CSAM things. We don't want things that, that give any, any person in the Department of War the ability to do things that we wouldn't want, whether it's chem, bio, those sorts of things. So guardrails. Are something that are negotiable based on what, what they are with all the companies. And they have different views on that when it comes to a constitution and how they bake it into their model. If you wait, if you look at the way Anthropic talks about how they've done that, they say everything refers back to that constitution and it is different. I think as of January 1st, 2026, it said that it preferred non-Western values in their constitution. So it It's wildly different from a set of principles on what they're trying to prove. But more importantly, when they deploy on our networks, they're deploying models that are tuned for national security purposes. And that's why the partnership with the executive team and the management is so important, because things are evolving in the threat landscape. And whatever guardrails, whatever principles they want to develop against has to be consistent with our values, our mandate, our restrictions even. And that's where the guardrails come in.

Emil, can I ask you, there's been a lot of reporting around Anthropic meeting in the White House with Suzy Wiles and others. That reporting has suggested that the White House, the administration, and Anthropic are trying to kind of tamp down the tensions, cool things off a bit, and maybe get to a point where the government can work very well with Anthropic, potentially even your department as well. Is that the case? Are you in a situation where you are looking at this differently, maybe finding a new way to move forward? With Anthropic?

I think, you know, I defer that particular question, Whitehouse. But I will say is that the cyber capabilities of all of these models is going to be this year's challenge for any company that has software deployed, really, or any consumer that has software on their laptop or on their phone. And that's because these models are starting to train on code. And Anthropic was the first out of the gate with Claude Code. But then there's Codex for OpenAI and there's coding tools forever. So they become adept at learning how to find vulnerabilities in code. So I think when, when we think about the first drop being Mythos, but there's going to be other drops. So I think the government's looking for how to work with all of these companies in the coming year to make sure that their capabilities are understood by us first. so that we can fix the, you know, any kind of issues that we have in, in the private and the public sector.

Emil, how concerned are you? And maybe you're using different compute, but there does seem to be real concern about a lack of compute power for some of these models, especially the sort of closed models, Mythos being the prime example. But this very idea of potentially, you know, you get to a period of time where there's not enough compute either for the government to be operating at the same time that either corporations are operating. I mean, it depends if you're operating on your own, your own cloud or something else. But how big a deal do you think that is and how should we think about that?

I think that, you know, we go back and forth as, as, as viewers here as to how are we overbuilding compute with all these data center buildouts? Are we underbuilding? And I think where we're at now as a, as a nation is that we need more compute overall. And that's why you see the rush to do all these deals, to stand up all these data centers. And I think that'll increase. The other thing to note is some of these models are changing what underlying infrastructure they work best on, whether it's a Trainium chip, whether it's AWS, whether it's Nvidia's newest. So, so things are moving and that's causing also sort of a little bit less uniformity, if you will. If it was more uniform, I think we might have close to enough compute. But I think you're going to see compute explode. If you look at the CapEx numbers announced by these hyperscalers, they're enormous. So that would indicate to me that sort of the nation needs more compute and the government could always use more compute as we do more things at the Department of War and other agencies do too.