The White House wants a bigger AI policy footprint just as the Pentagon’s guardrail fight gets ugly. Convenient timing—or a very revealing one. Quick catch-up before we dig in: Anthropic’s been pushing for binding controls on frontier AI while the White House builds a framework for government review of advanced models. The administration has briefed major labs on the process but reportedly kept the framework out of public view. So the fight is over who gets access, what gets disclosed, and whether labs get a say before the rules harden. This is Anthropic Pentagon Watch. Today: whether “national-security testing” brings actual limits—or gives Washington a cleaner label for squeezing labs into line. Hugo Lowell at WIRED has the details. WIRED says the White House is widening this AI policy push to open models. So while Anthropic gets squeezed over use limits, that same administration is building a bigger national-security testing tent around Anthropic and OpenAI. And that changes the closed-door White House framework: the policy push is widening to open models. A voluntary review process matters a lot more if it may decide which models get waved into national-security channels—and the public still can’t inspect the terms. Exactly. “National-security testing” sounds soothing until somebody has to define the test, the data access, and the permitted use. Vague policy language is where surveillance exceptions go to breed. It’s an escalation, not a finished rulebook. WIRED’s reporting signals broader White House involvement; it doesn’t tell us whether labs’ public contract carve-outs will survive inside that framework. How does a fight over what Claude can be used for become a Pentagon finding that Anthropic itself is a supply-chain risk? That sounds less like a contract dispute and more like a government-wide punishment. The bridge is a Pentagon law meant to protect military systems from supply-chain threats. Reuters says it hasn’t previously been used this way against a U.S. AI company. The dispute was over Anthropic’s limits on military use of Claude: reporting says the company wouldn’t allow mass surveillance of Americans or use of its technology to power fully autonomous weapons. The Pentagon wanted looser or unrestricted access. Before the designation, it was already asking major defense contractors whether they relied on Anthropic’s services, including Lockheed Martin; Reuters also reported contact with Boeing. Then, according to AP, Defense Secretary Pete Hegseth terminated Anthropic’s Pentagon and other government-agency work and applied the supply-chain-risk label. Axios says the practical consequence is that companies wanting to do business with the U.S. military could have to cut ties with Anthropic. That takes this from one direct customer relationship to a question for contractors across the defense supply base. Anthropic has challenged the move in court, arguing, Reuters reports, that the blacklisting violates free-speech and due-process rights. So is the designation already a settled finding that Anthropic is dangerous, or is it still essentially the Pentagon’s legal theory being tested in court? It’s very much contested. Anthropic filed complaints in California and Washington, D.C., after the designation, and Reuters says the case will test how far this obscure supply-chain law can reach. Legal experts told Reuters Anthropic appears to have a strong case. What to watch is whether a court limits that authority—and, in the meantime, whether contractors can or will disentangle Anthropic tools from government-related work. Here's OpenAI:
This language makes explicit that our tools will not be used to conduct domestic surveillance of U.S. persons, including through the procurement or use of commercially acquired personal or identifiable information. The Department also affirmed that our services will not be used by Department of War intelligence agencies like the NSA. Any services to those agencies would require a new agreement.
OpenAI’s March 2 addendum draws an actual line: no domestic surveillance of U.S. persons, including bought personal data, and no NSA use without a whole new agreement. That’s what a redline looks like when somebody bothers to write it down. That White House expansion we just covered is a useful pressure test. National-security testing can’t just be a roomful of officials nodding at “oversight”; the terms need to say whether these carve-outs travel with the model. OpenAI also says its deal has more guardrails than Anthropic’s. Fine—contract language with a competitive-comparison section bolted on. But now the Pentagon has a peer-lab benchmark it can’t pretend doesn’t exist. Hacker News, weighing in:
The Department of War may use the AI System for all lawful purposes, consistent with applicable law, operational requirements, and well-established safety and oversight protocols. The AI System will not be used to independently direct autonomous weapons in any case where law, regulation, or Department policy requires human control, nor will it be used to assume other high-stakes decisions that require approval by a human decisionmaker under the same authorities. Per DoD Directive 3000.09…
The human-control clause matters, but notice the limit: it applies where law, regulation, or Department policy already requires human control. It doesn’t create a universal ban on autonomous-weapons use. Here's one from Hacker News:
Not great? Seems kind of loose language? It isn't OpenAI saying no autonomous weapons use, but only that use must be consistent with laws, regulations, and department policies: "The Department of War may use the AI System for all lawful purposes, consistent with applicable law, operational requirements, and well-established safety and oversight protocols. The AI System will not be used to independently direct autonomous weapons in any case where law, regulation, or Department policy requires…
Exactly. “All lawful purposes” leaves the Department room to loosen its own policy and call the next use lawful. If the same institution can loosen the policy, that guardrail is pretty flimsy. Hacker News, weighing in:
For intelligence activities, any handling of private information will comply with the Fourth Amendment, the National Security Act of 1947 and the Foreign Intelligence and Surveillance Act of 1978, Executive Order 12333, and applicable DoD directives requiring a defined foreign intelligence purpose. The AI System shall not be used for unconstrained monitoring of U.S. persons’ private information as consistent with these authorities. The system shall also not be used for domestic…
There’s a meaningful distinction here: OpenAI’s public language separately bars domestic surveillance and sends NSA services into a new agreement. The intelligence terms still lean on FISA, the Fourth Amendment, and executive rules—so it will come down to enforcement. Have feedback, story ideas, or a correction? Send us a note at anthropicpentagonwatch at lantern podcasts dot com. We’d love to hear from you.
Links to every story we covered today are in the show notes, so take a look if you want to dig in. Thanks for listening, and we’ll be back with more tomorrow. That’s Anthropic Pentagon Watch for today. This is a Lantern Podcast.