For a week this was a fight over a label. Today it's a Council on Foreign Relations memo grading the whole policy setup — which means the grown-ups in the room finally showed up to read NSPM-11. Welcome to Anthropic Pentagon Watch. I'm Sarah, here with Devin — and today we're stepping back from the courtroom to ask what these rules actually do. And CFR's verdict on that framework? The word they keep reaching for is 'unresolved.' So let's find out exactly which unresolved they mean. Plus a Step Back segment that finally answers the rip-it-out question — what happens to agencies already running Claude. Let's get into the CFR read first. So CFR's June 9 piece is the first outside institution really grading NSPM-11 on substance. And what matters is what they leave open — because that's exactly where the Anthropic fight lives. Right, but here's my problem. NSPM-11 requires agencies to work with more than one AI provider. Meanwhile the Pentagon's in court fighting to keep Anthropic blacklisted — while leaning on Claude for damn near everything. Does CFR flag that contradiction or paper over it? That's the live tension. NSPM-11 pushes agencies to accelerate adoption. The injunction is still blocking the designation. A contracting officer reading both documents today sees two signals from the top, and they point in opposite directions. And I want to know what CFR doesn't say. If their experts walk through what's unresolved and never touch offensive cyber or the autonomous-weapons guardrails the Pentagon's been trying to strip — that silence is the answer. The separation-of-powers piece is what gets me. The people writing the rules for military AI aren't in Congress — they're in a contract dispute. CFR analyzing the architecture instead of the litigation is the first sourced data point outside that courtroom. So which part does CFR call 'right'? The 30-day model review window? The multi-provider mandate? Because if they're praising multi-provider as a feature while the Pentagon's whole posture is single-provider lock-in, somebody's grading on a curve. This brings us to the Step Back. The question all week has been: if Anthropic's kept off-limits, what happens to projects already running Claude? Short answer: you don't rip it out. You either get waivers, or you freeze expansion in place. So the blacklist creates a slow bleed, not an off-switch. And once waivers enter the picture, somebody runs that process — and somebody profits from how long it drags. The mechanism depends entirely on the contract pathway. Existing obligations are sticky. New work is where the freeze actually bites. And sticky cuts the other way too. The NSA signals-intelligence carve-out everyone treated as theoretical? If existing contracts can't be unwound, inertia locks that exception in. The hole in the red line starts to become the operating rule. So even the consensus institution calls the framework unresolved, and underneath it, enforcement looks like administrative paralysis. The conversation finally caught up to the mess. Read the CFR piece — it's the closest thing to a referee we've got. Here's Vinh Nguyen; Michael C. Horowitz at Council on Foreign Relations:
Notably, the memorandum requires agency’s to terminate contracts with AI companies that repeatedly limit government use of their technology (largely considered a response to the Pentagon’s legal battle with AI firm Anthropic), orders the Defense Department to update its policy on autonomous weapons, and vests accountability for AI use within the military chain of command rather than external regulators.
Yesterday, CFR's Vinh Nguyen and Michael Horowitz put a grade on NSPM-11 — and for once, the assessment is about the architecture, not the courtroom. When they say 'unresolved,' they're pointing at the gap we've been on all week: the memo tells agencies to terminate contracts with companies that limit government use, but the designation mechanism is still enjoined. So a contracting officer reading both today sees two executive signals pointing in opposite directions. Notice who Nguyen is — NSA's first chief responsible AI officer. So the man telling us how Washington learns to trust the AI it can't build spent his last job inside Fort Meade. That's the lens on this whole 'unresolved' verdict. And CFR confirms the memo orders DoD to update its autonomous-weapons policy and vests accountability in the chain of command — not external regulators. So when they say the memo gets things 'right,' I want to know whether 'right' includes pulling the guardrails inside the building that wants them gone. That's the separation-of-powers problem, now with an outside voice behind it. The rules for military AI are being written through a contract dispute and a presidential memo, while Congress mostly watches from the sidelines. CFR just confirmed the problem. Step back for me: if Anthropic is kept off-limits for federal agencies, what actually happens to existing government projects or defense contractors already using Claude — do they have to rip it out, get waivers, or just stop expanding it? So the short answer is: it's complicated, and 'just rip it out' is much harder than the order makes it sound. On February 27th, President Trump directed every federal agency to immediately stop using Anthropic technology. Defense Secretary Hegseth separately designated Anthropic a supply-chain risk to national security; that formal Pentagon designation came down on March 3rd, per Reuters. That label matters because, as Mayer Brown's legal team flagged for contractors, it doesn't just hit the Pentagon directly. It also reaches the many private contractors using Claude on government prime contracts and subcontracts, which means they may soon be barred from using it to perform that work too. The operational problem is that Claude is already embedded in classified networks. Scientific American reported that the Pentagon formally gave itself a six-month window to phase Claude out of those systems, and Nextgov's sources said it could take three months or longer for the military to regain access to a comparably capable tool on the classified side. Pentagon staffers, former officials, and IT contractors told Reuters they're genuinely reluctant to comply because they see Claude as better than the available alternatives. So the likely path is a phased, contested wind-down, with real capability gaps in the middle. If contractors are potentially on the hook too, are they waiting for formal contract language to change before they act, or does the supply-chain designation alone already bind them? That's exactly the live legal question right now — Goodwin's contractor advisory flagged that defense and civilian contractors should be actively assessing whether current or future activities involving Claude could put them in violation, even before new contract clauses are formally issued. So the prudent read from outside counsel is: don't wait for the paperwork. The thing to watch is whether the Pentagon issues formal acquisition guidance that explicitly incorporates the supply-chain designation into contract requirements, because that's what would give the ban real teeth across the contractor ecosystem. Got a tip, a story idea, or a correction for us? Send it to anthropicpentagonwatch at lantern podcasts dot com. We read every note, and your feedback helps make the briefing sharper.
If you want to dig further into any of today’s stories, we’ve put the relevant links in the show notes. Take a look there for the pieces and source material that caught your ear.
That’s Anthropic Pentagon Watch for this Wednesday. This is a Lantern Podcast.