A red line deadline for Anthropic

TL;DR: The clock is now ticking on the dustup between Anthropic and the Pentagon. Defense Secretary Pete Hegseth gave the company until Friday at 5:01pm to meet its demands during a reportedly “tense” face-to-face yesterday, a senior defense official confirmed to Tech Brew. The government is floating drastic legal measures to force Anthropic to roll back certain AI guardrails, but the company is said to be digging in its heels—despite rolling back some of its protections separately, if they affect its competitiveness. The next couple days might show just how resilient AI ethics pledges are in the face of potential business losses and DC pariah status. It’s also a test of just how irreplaceable Anthropic’s AI is.

What happened: An Anthropic spokesperson said that yesterday’s Pentagon meeting was a “good-faith conversation,” but a defense official’s description to Axios—“not warm and fuzzy at all”—sounds a bit more on point.

Hegseth reiterated to Anthropic CEO Dario Amodei that the department will declare Anthropic a “supply chain risk” and also invoke the Defense Production Act to force the company to allow use on the Pentagon’s terms—two different steps that would seem to contradict one another (OpenAI, Google, and xAI have all accepted the terms). Anthropic’s Claude was the only AI model approved for use in classified systems. As of Monday, the government has granted Elon Musk’s xAI (and its controversial Grok model) that status and is close to doing so with other AI companies it’s in talks with, the senior defense official told us (OpenAI has already been ramping up work with the military).

Anthropic’s leverage: Claude holds some cards that Grok doesn’t have, though: It’s generally considered to be better. And a senior administration official told Axios that other models “are just behind” when it comes to specialized government use cases.

Anthropic has positioned itself well for national security, and its models perform well on the use cases the Pentagon needs, according to Owen Daniels, associate director of analysis at Georgetown’s Center for Security and Emerging Technology.

And it may be hard to remove Claude depending on how deeply it’s already embedded, Chirag Mehta, vice president and principal analyst at Constellation Research told us. Anthropic’s relationship with Palantir also makes Claude tricky to disentangle.

“Swapping models is easy on paper but hard in production,” Mehta said. “The real work sits in the integration, the evaluations, and the accreditation trail.”

Why it matters: “Using the DPA to try to strongarm a private company to do what they want just strikes me as unprecedented and uncalled for,” Kristian Stout, director of innovation policy at the International Center for Law and Economics, told us. “We're not at war, we're not at threat of war. There's no asteroid landing that we need sudden emergency help with.”

The government also doesn’t usually label companies as “supply chain risks” unless they’re doing something like working with a foreign adversary. But Daniels said it is also “very uncommon” for a government contractor to draw boundaries the way Anthropic is.

“Part of the challenge here stems from the clash of cultures of these two organizations,” Daniels said. “From the Pentagon's perspective, there's […] the precedent that this could set in terms of companies trying to dictate terms to DoD.”

What comes next: Anthropic has until 5:01pm on Friday to comply with the Pentagon’s demands. But there’s always a chance the deadline could be kicked down the road. And Anthropic does hold some leverage here—the Pentagon clearly wants to keep using its models. But it’s also out on a limb without its industry peers backing up its stance. “They're really stepping out on their own in this regard,” Daniels said. —PK