WATCH: In SASC Hearing, Kelly Points Out Contradiction in Pentagon’s Decision to Brand Anthropic a Risk While Using Its AI for Operations

"There seems to be a theme here that anybody who questions this administration, questions the Secretary of Defense, offers some other route, some other idea, that isn't consistent with the leadership of the department, they're branded."

This week, during a Senate Armed Services Committee (SASC) Emerging Threats and Capabilities (ETC) Subcommittee hearing, Arizona Senator and Navy combat veteran Mark Kelly (D-AZ) demanded answers from Under Secretary of Defense for Research and Engineering Emil Michael on the Pentagon's decision to designate American AI company Anthropic a national security risk while U.S. troops were actively using Anthropic's Claude in combat operations against Iran.

Kelly laid out the contradiction directly: "If Anthropic's product is genuinely what Section 3252 contemplates a 'national security supply chain risk capable of sabotage, subversion or a mid-fight refusal that gets Americans killed,' how does your Department justify leaving that product in the hands of our troops in an active conflict one second longer?"

Kelly also warned that the Pentagon's treatment of Anthropic sends a chilling message to the AI industry: "There seems to be a theme here that anybody who questions this administration, questions the Secretary of Defense, offers some other route, some other idea, that isn't consistent with the leadership of the department, they're branded.

"After a lot of experience in the military and at other government agencies, if you're not asking the hard questions, especially uphill, you're not getting the best out of your people and you're not getting the best decisions. And my concern here is you, as you branded a company in a way that's harmful for them and their employees, thousands of people that work there. But you also may have ultimately put our servicemembers at risk if they are, in fact, the best company to provide these services."

Sen. Kelly questions Under Secretary Michael during a SASC hearing.

Click here to download a video of Kelly's remarks. See the transcript below:

Senator Kelly:

Mr. Michael, under Title 10 U.S. Code Section 3252, the statute that your Department invoked against Anthropic as a supply chain risk, it's defined as the risk that an adversary may, and this is a quote, "sabotage, maliciously introduce unwanted function, or otherwise subvert," end quote, national security systems. That's the language that Congress chose. "Sabotage or subversion." Language that implies the undermining of our national security. On an "All In" podcast in March, you explained this in your own words, saying, this is another quote here, "What if this software went down, some guardrail, kicked up, some refusal happened for the next fight like this one, and we left our people at risk?" So that was the danger that you described. A mid-mission failure that would put American troops lives in danger. Now here are the facts. As we sit here today, Anthropic has been designated a supply chain risk. And your Secretary Pete Hegseth called the decision "final." The President has ordered every federal agency to stop using their products. And yet, U.S. service members have used Anthropic's Claude in combat operations against Iran since the strikes began on February 28^th. Your Secretary's 180-day phaseout has not expired as of March 24^th. Your own Chief Information Officer, Kirsten Davies, testified to the Senate that Claude's use in U.S. military operations in Iran was, quote, "active right now." So, I've got a question. It's two parts. Part one, if you are correct, if Anthropic product is genuinely what Section 3252 contemplates a "national security supply chain risk capable of sabotage, subversion or a mid-fight refusal that gets Americans killed," how does your Department justify leaving that product in the hands of our troops in an active conflict one second longer?

Under Secretary Emil G. Michael:

When they were declared a supply chain risk, we stopped all updates permitted from that company. So, we froze the code and the cloud provider that it was hosted on, froze any access so that no change could be made in the code so there was no risk of any mid-battle change in the operations of that.

Kelly:

But that also means there is no risk in the code as it was on the day that you said it could be used against the troops in a way that could put them at risk.

Michael:

No, Senator, because, before you came in, I listed the things that were against their terms of service, that, for example, you were not allowed, by their terms of service that we live under today, to do battlefield management applications, target any military infrastructure of adversaries, direct interceptions, like of a drone, or develop weapons systems. And that could still be built into the code. So, we just avoiding the use cases.

Kelly:

Let me ask you a second question though. If Anthropic's product, it seems like what they're using is safe enough now to remain in our soldiers' hands in a shooting war against a foreign adversary, how does the department justify branding an American company a "sabotage and subversion risk" under a statute that Congress wrote for foreign spies? Seems a little bit extreme. And one of these two things has to be wrong. Which is it?

Michael:

The sophistication with which we use AI at the Department of War is very nascent right now. The way we use it is still very early days. What we're worried about with the terms of service that they had and their posture towards the Department, which when they questioned the Maduro raid and whether their software was used inappropriately, gave us the sense that this was not a reliable partner to deal with, given their-

Kelly:

Because they asked questions about how their product was being used? It means they're not reliable?

Michael:

That, in conjunction with their written terms of service which prevent the use cases that we would like to advance into battlefield management, directing interceptions, developing weapons systems.

Kelly:

Well, here's the thing. There seems to be a theme here that anybody who questions this administration, questions the Secretary of Defense, offers some other route, some other idea, that isn't consistent with the leadership of the department, they're branded. In this case, this company is branded a "subversion risk" and essentially fired because they started trying to have a discussion about, "Hey, what's the what's the best use of this?" Or "Why did you do this this way or that way?" I think-My view on this after a lot of experience in the military and at other government agencies, if you're not asking the hard questions, especially uphill, you're not getting the best out of your people and you're not getting the best decisions. And my concern here is you, as you branded a company in a way that's harmful for them and their employees, thousands of people that work there. But you also may have ultimately put our servicemembers at risk if they are, in fact, the best company to provide these services. They seem to be. Thank you, Madam Chair.

Michael:

Can I?

Chair Enrst:

Yes, and you may, because, Secretary Michael had listed now the partners that we do have that will be filling this gap. But yes, I'd like you to respond.

Michael:

Senator, every other major AI company, every single one agreed to our terms. So, this wasn't about questioning whether we were-they dared to question us. This is, we wanted to use their products for all lawful use cases. And Microsoft agreed to that. OpenAI agreed to that. NVIDIA agreed to that. Google agreed to that. So, every other company-

Kelly:

How many of those were after Claude essentially got fired?

Michael:

They were all after because we didn't-we only had one provider before we got in the seat. When we looked at the contract of that provider, we said, "We can't grow with that provider."

Kelly:

Do you think there's any chance they saw what happened to Claude, saw an opportunity, and then said, "Well, I know what I have to agree to?"

Michael:

Absolutely not. Maybe OpenAI. I'll retract and say maybe one company that sort of tried to slide in one of the direct competitors. But Google has been a long-time partner of the Department. Microsoft, NVIDIA, real big companies with proper corporate governance, went through their legal teams and agreed to our terms of all lawful use cases where Anthropic would not. So, that should say something that our terms weren't unreasonable. We weren't trying to be unreasonable vis-à-vis any one partner for any specific reason that they didn't demonstrate themselves.

Kelly:

Who do you think is best equipped to provide these services today?

Michael:

It depends on the application. I'll give you some examples as to why. Google, for example, has a lot of video imagery because of their YouTube and Nest cams. And therefore, from a robotic standpoint, they may provide better capability from autonomous weapons standpoint. NVIDIA has an open-source product, which is going to be a crowdsourced capability that can compete with the Chinese open-source models, which are currently infiltrating a lot of the U.S. companies. So, each of them have different capabilities, which is frankly why, independent of Anthropic, I would have ensured that we had multiple providers because the leaderboards change on these providers every few months. It wouldn't surprise me if next month you have OpenAI, have a better cyber model than Mythos. So, for us to be responsible to the warfighter, I've got to have all the options possible as these models change in advance.

Kelly:

What do you do about Mythos now? Because that's Claude.

Ernst:

I think we'll start wrapping up. Yes. Thank you.

Michael:

I think Mythos is the first cyber weapon model. OpenAI will have a cyber weapon model. Gemini will all this calendar year, and then next year we'll be worrying about bio and chem models as they are to train them with science and physics and equations and math. And we're going to have an ongoing discussion, probably, and hopefully with Congress, about how we govern the rollout of these things that have good effects, maybe they can cure diseases, but then they have bad effects when in the hands of bad actors. And that's something that we're going to have to deal with as a nation.