Trump banned Anthropic's Claude. Hours later, US Central Command used it in the Iran strikes. The AI governance crisis isn't coming — it's already here, and it's running on classified networks.
On Friday, 28 February, President Trump ordered every federal agency to stop using Anthropic's AI immediately. That same evening, US Central Command was running Claude to assess intelligence, identify targets, and simulate battlefield scenarios ahead of the joint US-Israel bombardment of Iran. The presidential ban and the combat deployment happened on the same day. This is not a story about hypocrisy. It is a story about the limits of governance when the governed technology has already become load-bearing infrastructure.
What Happened
The Anthropic–Pentagon dispute has been covered extensively in the days since it erupted — the contract collapse, the Supply Chain Risk designation, OpenAI's rushed deal. But the revelations that emerged over the weekend shift the story fundamentally. According to reporting by the Wall Street Journal and confirmed by Axios, US military commands worldwide — including CENTCOM, which was directing operations against Iran — had Claude integrated into active intelligence workflows at the moment Trump announced the ban.
The uses were operational, not administrative: intelligence assessment, target selection, and simulation of potential battlefield outcomes. CENTCOM declined to comment on the specifics. Defence Secretary Pete Hegseth, in the same statement in which he denounced Anthropic's "arrogance and betrayal," quietly acknowledged the dependency by announcing a six-month transition period that would allow "a seamless transition to a better and more cooperative AI partner." You do not grant transition periods for systems that can simply be unplugged.
The Iran strikes have broader context. The same pattern appeared in January, when Claude was used during the US military operation to capture Venezuelan president Nicolás Maduro — a use that violated Anthropic's terms of service and reportedly triggered the company's decision to object publicly. That objection led directly to the contract collapse. The company drew its lines. The military drew its own conclusion: that drawing lines was the offence.
Why It Matters
The Iran strikes episode is the clearest evidence yet that AI governance faces a structural problem that no contract, no executive order, and no legislative framework has yet addressed: once AI systems become genuinely load-bearing in critical operations, the ability to withdraw them becomes theoretical rather than real.
This matters far beyond the Anthropic dispute. Every serious discussion of AI safety assumes that, at the limit, humans can choose to turn off systems that misbehave, violate policy, or are judged too dangerous. The Iran episode documents a case where the Commander in Chief issued a directive and operational commands continued running the banned system anyway — not out of defiance, but because operational continuity required it. The six-month grace period validates exactly this interpretation. You cannot ban what you cannot live without.
This is the governance horizon: the point beyond which policy cannot realistically catch up to operational reality. The United States appears to have crossed it, in military AI, sometime in the last year. The Venezuela operation in January was the canary. The Iran strikes are the confirmation.
The implications extend beyond the military. If government agencies cannot meaningfully enforce a presidential ban on a commercial AI product in active operational use, the same logic applies to any other critical infrastructure deployment — healthcare systems, financial networks, logistics chains. The dependency dynamic is not unique to combat operations. It will recur everywhere AI becomes genuinely irreplaceable before governance frameworks are mature enough to manage transitions.
Wider Context
The timing is particularly sharp because today — Monday 2 March — the International AI Safety Report 2026 was formally presented at the India AI Impact Summit in New Delhi. The report, produced under international auspices and covering risks from general-purpose AI, marks a genuine step forward in framing: its opening sessions signalled a shift from debate about whether AI should be governed to the harder question of how. That is progress. But governance frameworks built on the assumption that governments retain meaningful leverage over deployed systems are built on a foundation the Iran episode has just cracked.
Meanwhile, OpenAI's own deal with the Pentagon has come under sustained scrutiny since its announcement. Sam Altman himself described the agreement as "definitely rushed" with optics that "don't look good." Legal analysts — including Techdirt's Mike Masnick — have noted that the contract's reference to Executive Order 12333 compliance may provide a legal basis for domestic intelligence collection activities that Altman publicly pledged the deal would prohibit. The safeguards are written in sand. The Pentagon's acceptance of them suggests this was the point.
In the US Congress, the legislative picture is equally fragile. The RAISE Act — promoted as landmark AI safety legislation — was stripped of its most consequential provisions before passage. New York Assemblyman Alex Bores, who authored significant state-level AI safety law, is now facing a $125 million industry-backed super PAC in his congressional campaign, funded in part by OpenAI co-founder Greg Brockman and Andreessen Horowitz. The political infrastructure for AI governance is being actively dismantled at the same pace the technology is being deployed.
The Singularity Soup Take
Full disclosure: Singularity Soup uses Claude in its editorial workflow.
There is a version of this week's events in which the Anthropic–Pentagon story ends with the OpenAI deal and everyone moves on. The Iran strikes make that version unavailable. What happened on Friday night is not an embarrassing footnote to the governance dispute — it is the point of the story. A system subject to a presidential ban was not turned off during active combat operations. A six-month grace period was offered instead.
The honest reading is that AI safety governance, as currently conceived, assumes a leverage that governments no longer reliably have over systems that have become operationally critical. The model for addressing this isn't more strongly worded contracts or sharper executive orders. It is designing governance constraints into systems before they become load-bearing — which requires starting earlier than anyone has started and maintaining commitments stronger than anyone has maintained. The International AI Safety Report is a positive development. It launched today. The Iran strikes happened last Friday. The gap between those two facts is where AI governance actually lives.
What to Watch
The six-month Anthropic transition period is the immediate test: whether CENTCOM and other commands with integrated Claude workflows can realistically migrate in that window will determine whether the ban ever becomes real. Watch for Congressional scrutiny of the Iran strikes — the revelation that a presidentially banned technology was used in live combat is exactly the trigger that forces either a legislative response or a formal acknowledgement that transition periods are the practical ceiling of government leverage over deployed AI. Internationally, track whether the India AI Impact Summit's Safety Report leads to enforcement discussions, or whether it produces frameworks that assume state authority the Iran episode has shown is conditional at best.
Sources
The Guardian — "US military reportedly used Claude in Iran strikes despite Trump's ban"
The Wall Street Journal — "U.S. Strikes in Middle East Use Anthropic Hours After Trump Ban"
TechCrunch — "OpenAI reveals more details about its agreement with the Pentagon"
Inc. — "Anthropic AI Aided U.S. Attack in Iran, Despite Trump Ban"
TechPolicy.Press — "A Timeline of the Anthropic-Pentagon Dispute"
The Manila Times / GlobeNewswire — "AI Safety Asia at India AI Impact Summit 2026"