Pentagon AI deals are testing Claude’s red lines—and it looks like somebody else may be holding the ruler. Here’s how we got here: Anthropic’s push for binding frontier-AI controls has collided with the White House effort to review advanced models before release. OpenAI, Anthropic, Google, and other tech companies had already been in Washington as the administration weighed cybersecurity-focused reviews. Open-source models were still outside the process. That left companies and watchdogs waiting to see how government review would actually work. This is Anthropic Pentagon Watch. Today: a giant military data deal, a rival AI vendor cleared for sensitive work, and frontier-model rules being discussed behind closed doors. First, the money. From Brandi Vincent at DefenseScoop:
But in response to repeated requests for more information in recent weeks, a Pentagon official told DefenseScoop on Friday that GSA, on behalf of the CDAO, formally awarded the contract to AFS on June 25. The single-award task order has a potential ceiling value of $821,270,264 over a five-year period of performance, including a one-year base period and four one-year option periods.
Eight hundred twenty-one million dollars, awarded June 25—and DefenseScoop had to keep asking before the Pentagon confirmed the ceiling or the five-year timeline. This is the data plumbing for operational AI, and they treated the purchase order like classified poetry. The structure matters: one Accenture Federal Services task order, a one-year base, then four annual options through GSA’s Alliant 2 vehicle. The Pentagon can call it competitive, but it still gives one integrator enormous leverage over what data is usable downstream. And War Data Platform is just Advana with a new label. A former senior defense official told DefenseScoop that the same types of firms, and many of the same characters, helped create the old problems—now they’ve got an $821 million fix-it contract. Rebranding is cheaper than remediation. This award runs for five years, so contractors and civil-liberties lawyers now have a concrete architecture to scrutinize. From Daniel Terrill at Military Times:
In Wednesday’s announcements, the company explained that its Missionforce National Security platform received Impact Level 5, or IL5, authorization, meaning it is “trusted” to address the “most sensitive, unclassified missions.”
“Admin tasks” is a very cozy label for autonomous agents handling controlled unclassified information and National Security System data. Once Missionforce is making decisions in that pipeline, somebody has to draw the line before deployment—not afterward. Salesforce got IL5 authorization for what Military Times calls the most sensitive unclassified missions, on top of a $5.6 billion Missionforce contract from January. That’s a serious procurement lane. It goes well beyond a pilot with a clever name. And Salesforce is selling the jump from chatbots to “autonomous AI execution” as the milestone. Great—so the agents aren’t just drafting the memo anymore. They’re part of the machinery that moves it. IL5 tells us how sensitive the data is; it doesn’t tell us the operational boundary. The Pentagon and Army Human Resources Command have approved a platform built on commercial Agentforce technology. The oversight question is what decisions those agents may actually initiate. The Hill, with Julia Shapero and Miranda Nazzaro:
The Trump administration’s decision to keep its new framework on government AI testing out of public view has blindsided the tech policy community, who waited months to see the controversial process ironed out. The framework was expected to detail the process and benchmarks for government testing of private AI models ahead of their public release.
This has gone from private briefings to a transparency fight. The Hill reports that attendees were told the AI-review framework will not go public. Anthropic got a seat Tuesday; everyone else gets a federal release process they can’t read. The room included Anthropic, Meta, OpenAI, and Microsoft—and, per Forbes, Nvidia and some smaller companies. Government gets early access to models, handpicks “trusted partners,” then tells the public to admire the mystery box. Classified cyber benchmarks may need classified handling. But the voluntary framework deciding who becomes a “covered frontier model” is a separate governance choice. Keeping that secret gives agencies enormous discretion with very little outside check. Especially while Anthropic is fighting over limits on surveillance and autonomous weapons. Join the closed-door review, accept terms you can’t publicly challenge, and good luck explaining where your contract boundaries went. The Pentagon says it wants AI for routine support and more sensitive military work. But when does that become surveillance or autonomous-weapons use—and who gets to draw that boundary? There isn’t one clear public rule. The fight is happening in the contract terms. Reuters reported that Anthropic sought safeguards against using its technology to target people or in autonomous weapons, while the Pentagon maintained it should be able to deploy the technology regardless of a company’s usage policies. The negotiations also covered AI use on classified Pentagon systems, according to The New York Times. Reuters reported that the department was pressing major AI firms to make tools available on classified networks without their standard restrictions. One flashpoint: the Pentagon wanted to use Claude to analyze bulk commercial data on Americans, MIT Technology Review reported, while Anthropic wanted to bar mass-surveillance use. A senior official told Reuters that broad operational restrictions could hinder real-time planning and combat operations. In practice, whoever sets the governing terms sets the boundary: Anthropic through its conditions of use, or the Pentagon through procurement and access to military systems. And if the Pentagon rejects those conditions, is that just a disagreement with one company—or can the government effectively force the issue across the AI industry? Yes—the procurement fallout is immediate. CBS reported that the Pentagon had awarded Anthropic and rivals including OpenAI, Google, and xAI $200 million contracts, and that Anthropic faced the prospect of losing government-contract access after failing to reach an agreement. Watch whether other AI suppliers accept less restrictive classified-network terms, or whether Anthropic’s limits become a model for future military AI contracts. If you’re following Anthropic and the Pentagon, check out AI Daily Briefing: top AI news for engineers, founders, and investors every weekday, with real capabilities versus demo hype explained fast. Find it wherever you listen to podcasts.
Links to every story we covered are in the show notes. If something caught your attention, take a look and read further at your own pace.
That’s Anthropic Pentagon Watch for today. This is a Lantern Podcast.