When US forces captured Venezuelan ex-president Nicolás Maduro, few outside the operation knew exactly what tools were used. Now, one of those tools may be at the center of a growing dispute between Anthropic and the Trump administration.

The Anthropic-Pentagon feud in short

Anthropic prohibits two specific uses of its products: mass surveillance of American citizens, and fully autonomous lethal weapons, ie systems that select and engage targets without a human making the final call. When Maduro was captured, Anthropic executives reached out to their counterparts at Palantir to ask whether Claude had been used in ways that crossed those lines.

That question didn’t go over well. Palantir flagged the inquiry to the Department of Defense, and given Anthropic CEO Dario Amodei’s well-known stance on AI safety and responsible use, the DoD apparently read it as Anthropic getting political about military operations. Tensions followed.

Was Claude used to kill anyone during the Maduro operation?

Almost certainly not in any direct sense. Claude did likely not autonomously select targets or trigger weapons. But was it used somewhere in the mission? Almost certainly yes. AI tools like Claude are typically used for things like synthesizing intelligence reports, summarizing signals or open-source data that has been downgraded to a shareable classification level, or supporting logistics and planning. In other words: staff work, not trigger-pulling.

The more interesting question

Here’s what’s actually worth understanding: how does the US military use a commercial AI product like Claude on a mission involving classified information?

Data related to an operation like this would likely be classified at CONFIDENTIAL or SECRET level, potentially with some material sitting at TOP SECRET. It’s easy to think that’s the kind of data you’d want to keep far way from AI vendors.

The reason Claude can be used at all on classified networks comes down to a specific accreditation called IL6. The Pentagon contracts with Palantir - a CIA-backed data analytics company co-founded by Peter Thiel - and Palantir runs Claude through AWS infrastructure that has earned the IL6 designation. That’s essentially the US government’s stamp of approval for processing SECRET-level information. That infrastructure sits on SIPRNet, the classified government network, and classified data never leaves that environment.

Getting to IL6 is not quick. It takes years of security audits, cleared personnel, strict physical facility requirements, and compliance with hundreds of DoD security controls. It’s a designation normally associated with the most trusted defense contractors — not AI companies that might push back on how their models are being used.

Which is part of why the Pentagon’s response carries some weight. The DoD is reportedly beginning to view Anthropic as a “supply chain risk” — a label that is usually reserved for foreign adversaries, not American AI labs.

What happens if the Pentagon bans Claude?

If the Pentagon follows through on banning it, military personnel working on classified networks could be without access to that capability for at least three months, possibly longer.

What’s confusing about this story as that there are reports saying that Claude is the only frontier model available to the Pentagon. However, both Microsoft’s Azure OpenAI Service and Google’s Gemini models also hold IL6 accreditation, so Claude is not the only path forward. But maybe Claude is the only model Palantir offers within certain designated AWS environments.

Conclusion

At the heart of this story is a tension that is only going to grow: AI companies have terms of service, and the military has missions. When those two things come into conflict, who blinks first? Anthropic asked a reasonable question about how its model was used. The Pentagon responded by questioning whether Anthropic can be trusted as a vendor at all. The answer to that question will shape which AI tools the US military has access to on classified networks for years to come.

That’s it, I hope you enjoyed it. Feel free to connect with me on LinkedIn.