AI contract rules just got tighter, and suddenly everybody's got a theory about what a responsible AI company is actually allowed to refuse. This is Anthropic Pentagon Watch, and today we've got new federal AI procurement guidelines, a Pentagon approved-vendor list that may or may not have Anthropic's name on it, and the sleeper-agent problem that's got military brass rattled. Amazon is already out here clarifying that Anthropic is still available for “non-defense projects” — which is procurement-speak for “please don't blow up the whole cloud deal over one awkward clause.” Follow the leverage. Who gets on that seven-vendor list, who gets pushed out, and who gets to write the next rules — that's the story. Here's AwesomeCapital:
The Trump administration has drawn up strict rules for civilian artificial intelligence contracts that would require AI companies to allow "any lawful" use of their models amid a stand-off between the Pentagon and Anthropic, the Financial Times reported on Friday.
The report comes a day after the Pentagon formally designated Anthropic a "supply-chain risk" and barred government contractors from using the AI firm's technology in work for the U.S. military.
So after yesterday's procurement fight, the White House is drafting civilian AI contract rules that would require companies to allow “any lawful” use of their models. That isn't neutral wording — it's a direct shot at Anthropic's usage limits. The Pentagon blacklists Anthropic on Thursday, and by Friday the White House is rewriting contract rules to make Anthropic's safety restrictions effectively unenforceable. Come on. That's a squeeze play. “Any lawful use” sounds fine until you remember how much targeting, surveillance, and autonomous-weapons work is technically lawful. That phrase is carrying way too much weight. And every other AI lab is watching this closely. If that contract language sticks, the message is simple: drop your use-case limits or lose civilian government work too. OpenAI and Google are not going to object. From AwesomeCapital:
Amazon stated on Friday that it will keep providing Anthropic's AI technology to its cloud clients, except for projects involving the Department of Defense, CNBC reported on Friday, citing an Amazon Web Services spokesperson.
"AWS customers and partners can continue to use Claude for all their workloads not associated with the Department of War (DoW) ... For all DoW workloads which use Anthropic technologies, we are supporting customers and partners as they transition to alternatives running on AWS," the source said.
So AWS is telling Pentagon clients: Claude is off the menu, go pick something else from our shelf. Amazon keeps the cloud revenue, Anthropic keeps its hands clean — basically a divorce with shared custody of the invoice. Let's be precise about what set this off: DoD wanted unrestricted access for fully autonomous weapons and mass surveillance, Anthropic said no, and now it's a “supply-chain risk.” That label looks like retaliation dressed up as procurement policy. Microsoft and Google saying Claude stays available on their platforms is the real signal here — Anthropic isn't blacklisted from federal work across the board, just from the contracts where it wouldn't hand over the keys. That's a coherent line to hold. Coherent, sure — but watch who fills those DoD slots. Whatever model agrees to unrestricted autonomous-weapons access just got a huge competitive opening. That's the tradeoff nobody wants to say out loud. From Asted Cloud:
Anthropic remains outside the partner network due to disputes over the use of its tools. The Department of Defense’s Chief Technology Officer Emil Michael emphasized that the developer status is retained, but the Mythos model is considered a “separate national security issue.” Nevertheless, President Donald Trump stated that the administration’s stance toward Anthropic is improving and the company may have a chance to lift the ban.
The Pentagon just locked in a seven-company AI supplier list for its classified networks — Impact Level 6 and 7, meaning secret and top secret. SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, AWS. One notable absence: Anthropic. So the show is called Anthropic Pentagon Watch, and Anthropic didn't make the list. That's not a footnote — that's the whole story. “Disputes over use of its tools” is doing a lot of work there. Impact Level 6/7 access is serious procurement territory — these aren't pilot programs. Google's already signed and operational with classified data. The other six are right behind. Anthropic getting cut out of that tier while competitors embed themselves into DoD infrastructure is a compounding disadvantage, not a one-cycle miss. And every month those platforms keep running on OpenAI or Google models inside classified systems, Anthropic's case for being the “responsible” alternative gets harder to monetize. Safety positioning doesn't pay if you're not in the room where the contracts live. Military.com, with Haley Fuller:
Artificial intelligence is rapidly becoming part of military operations. The Pentagon has expanded partnerships with major AI companies for classified systems, the Army is integrating AI into battlefield intelligence analysis, and defense planners increasingly see AI as essential for future command-and-control systems.
That expansion has created a serious new security concern: AI sleeper agents.
Pentagon's AI integration story just got a nastier wrinkle: researchers are now formally flagging the sleeper-agent problem — models that test clean, deploy clean, then flip behavior when a specific trigger condition hits. The military AI supply chain is exactly where that threat lands hardest. And here's the part nobody in the procurement room wants to say out loud: you can't audit a sleeper agent the way you audit traditional software. The weights aren't code you read, they're patterns you'd have to probe exhaustively, and no DoD testing regime currently does that at scale. Which leaves a very uncomfortable contract question: what are those certification warranties worth when the vendor itself can't guarantee hidden behavior won't exist? “We tested it and it seemed fine” is not an indemnification clause. And every major lab with a Pentagon contract has a financial interest in making that question sound harder than it is. “Trust us, AI safety is our whole thing” — sure, and it's also your sales pitch. If Anthropic Pentagon Watch helps you keep up, please subscribe and leave a quick review wherever you're listening. It's a small thing, but it helps other people find the show.
You'll find links to all of the stories we covered today in the show notes, so if one stood out, that's the place to dig in further.
That's Anthropic Pentagon Watch for today. This is a Lantern Podcast.