The Pentagon found a way to dress up contractor leverage as interoperability. This is Anthropic Pentagon Watch. Today, one lab gets classified access with guardrails, another gets squeezed for refusing an open-ended clause—and allied models may not even agree on the battlefield. So what, exactly, is the Pentagon willing to accept from an AI contractor? Let’s start with the contract language. OpenAI reportedly got access to a classified Pentagon network while Anthropic was threatened over its safeguards. So what did OpenAI actually agree to, and was it really a different deal? OpenAI’s public description says its February agreement lets its technology be used in classified settings, but bars intentional domestic surveillance of U.S. persons, including through commercially acquired personal or identifiable data. OpenAI also says the Defense Department affirmed that its services won’t be used by intelligence agencies such as the NSA unless there’s a separate agreement. MIT Technology Review reports that OpenAI said its agreement also preserved restrictions on autonomous weapons and mass domestic surveillance. The dispute with Anthropic was over the Pentagon’s demand that it permit “all lawful uses,” according to The New York Times. After Anthropic resisted, the department labeled it a supply-chain risk. Reuters had reported earlier that the Pentagon was pressing major AI companies to make tools available on classified networks without their standard restrictions. But if the Pentagon can call Anthropic a supply-chain risk after it says no, doesn’t that make those written guardrails pretty fragile for every other contractor? And that’s the leverage question. A classified-network deal can be a major procurement opportunity, while a supply-chain-risk designation can shut a company out of government use. AP later reported that the Pentagon made classified-system deals with seven companies, with Anthropic notably absent. Watch whether the Pentagon accepts company-specific limits in practice—or keeps treating unrestricted “lawful” use as the price of access. From James JB Park and Sanghyun Ahn at The Diplomat:
In 2025, the Center for Strategic and International Studies (CSIS) and Scale AI tested seven major foundation models across 400 diplomatic and security scenarios. Despite being presented with identical international crises, the recommendations from the seven models diverged significantly. Some were uniquely hawkish, and every model exhibited some degree of national bias. This demonstrates that AI is by no means a homogenous tool.
Seven models, 400 identical crisis scenarios—and sharply different advice. Some were uniquely hawkish; every one showed national bias. So when Washington talks about AI-enabled coalition warfare, whose machine gets to be the trigger-happy tie-breaker? The Diplomat’s point gets operationally awkward fast: interoperability isn’t just getting U.S., Japanese, and South Korean systems to exchange data. It’s whether commanders can act together when their decision-support tools read the same battlefield and recommend different moves. And it puts the Pentagon’s pressure campaign in a worse light. We just discussed OpenAI keeping limits on domestic surveillance and autonomous weapons in a classified-network deal; meanwhile, the department wanted Anthropic to accept “all lawful uses.” Before demanding fewer guardrails, maybe prove the coalition stack can tell bias from battlefield insight. Procurement loves a common interface. Common judgment is harder to buy. If allied models diverge on the same crisis, contract access without shared evaluation standards could turn interoperability into a very expensive confidence game. If this briefing’s useful, please subscribe and leave a review wherever you’re listening. Reviews help other people find Anthropic Pentagon Watch, and they help us keep bringing you the stories that matter.
Links to every story are in the show notes if you’d like to dig deeper. Check out the ones that caught your attention, and we’ll be back tomorrow. That’s Anthropic Pentagon Watch for today. This is a Lantern Podcast.