When Claude Told the Pentagon "No": Anthropic vs. the Department of War
When Claude Told the Pentagon "No":
Anthropic vs. the Department of War
An AI company refused to let the U.S. military use its model for autonomous weapons and mass surveillance. The government called it a national security threat. The same weekend, Claude was used in strikes against Iran. We look at the numbers — and what they actually mean.
On Friday, February 27, 2026, at 5:14 p.m. Eastern time, U.S. Secretary of Defense Pete Hegseth posted on X that he was directing the Department of War — formerly the Department of Defense, renamed by executive order in September 2025 — to designate Anthropic a "Supply-Chain Risk to National Security."
The same evening, President Trump posted on Truth Social directing every federal agency to "IMMEDIATELY CEASE" all use of Anthropic technology. In a Politico interview days later, he said he had "fired them like dogs."
That weekend, according to multiple news reports confirmed by Anthropic's own CEO, Claude was being used to support U.S. military strikes against Iran.
Let that sentence sit for a moment.
The government declared a company a national security threat on Friday. The same company's AI model was used in active combat operations on Saturday. This is not a headline from a dystopian novel. This happened — and it raises questions that are statistical, legal, and deeply philosophical all at once.
How We Got Here: Eleven Months in Sequence
What the Pentagon Actually Asked For
The dispute comes down to two sentences that Anthropic refused to remove from its contract. Understanding them is essential, because the political framing — "woke AI company refuses to help the military" — deliberately obscures what was actually at stake.
The Pentagon's counterargument, articulated by Undersecretary Emil Michael in a CBS News interview, was that existing federal law already prohibits both of these things — making Anthropic's contractual restrictions redundant. If the law already bans autonomous weapons and domestic mass surveillance, why does a private company need its own contractual veto on top of that?
Anthropic's answer is precise, and it is worth taking seriously as a matter of probabilistic reasoning: a law the government can change is not the same as a contract the company retains the right to enforce.
This is a classic argument from conditional probability. The probability that a safeguard holds is not just P(safeguard exists today). It is P(safeguard exists today) × P(safeguard survives future legal or executive modification). A contractual restriction held by a private company that has independent incentives to enforce it has a meaningfully different risk profile than a statutory restriction held by a government that can amend its own statutes.
Let P(misuse | law only) = probability of misuse given statutory protection alone.
Let P(misuse | law + contract) = probability of misuse given statutory + contractual protection.
Anthropic's argument is simply that P(misuse | law + contract) < P(misuse | law only), because the contract adds an independent enforcement layer. The Pentagon's argument is that the difference is negligible. This is an empirical disagreement, not a values disagreement — and the empirical evidence from history of government surveillance programs does not obviously favor the Pentagon's position.
Designated a Threat on Friday. Used in Combat on Saturday.
The most statistically remarkable fact in this entire episode is not the designation itself. It is the timing.
Hegseth's transition plan explicitly allows Anthropic to continue providing services to the military for six months — for "a seamless transition." The Wall Street Journal and CNBC both reported that U.S. strikes in Iran used Anthropic's technology in the days immediately following the ban announcement. Anthropic's CEO confirmed this himself, stating the company would continue providing models to the DoW at "nominal cost" for as long as necessary.
"The government cannot simultaneously claim a vendor poses an acute supply chain threat requiring emergency exclusion and that it's perfectly safe to keep using the vendor for active combat operations."
— Lawfare, March 2026From a logical consistency standpoint, the government's position contains a clear contradiction. A supply chain risk designation implies the vendor cannot be trusted in sensitive government systems. Continuing to use that vendor in live military operations implies the opposite. Both statements cannot be simultaneously true — at least not within any coherent definition of "supply chain risk."
The Lawfare analysis — written by a former Trump-era White House adviser — argues this contradiction may doom the government's legal position before the case even begins. Courts examining emergency national security designations generally give the executive branch substantial deference. But they also require internal consistency. The six-month transition period, combined with active operational use, suggests the urgency framing is pretextual.
By the Numbers: How Unusual Is This?
The statute Hegseth invoked — a combination of the Federal Acquisition Supply Chain Security Act (FASCSA) and 10 U.S.C. § 3252 — was designed to address adversarial foreign suppliers embedding vulnerabilities in government technology. Think Chinese telecoms. Think state-sponsored hardware backdoors. The legislative history of these statutes contains no contemplation of applying them to a domestic company over a contractual disagreement about terms of use.
Anthropic's legal counter-argument, which Dario Amodei telegraphed in his public statements, focuses on a statutory interpretation point: the designation "exists to protect the government rather than to punish a supplier," and the law explicitly requires the Secretary to use "the least restrictive means necessary." Simply terminating the contract — an option Anthropic said it supported — would have been a far less restrictive means. The choice to pursue a supply chain designation instead suggests a punitive motivation, which is legally relevant under the major questions doctrine.
The most extreme interpretation of Hegseth's X post — that all companies doing business with the DoW are barred from any commercial activity with Anthropic — would have forced AWS and Google Cloud to drop Anthropic as a customer. Since every major cloud provider is a DoW contractor, this would have effectively deplatformed Anthropic entirely, taking the company offline. Legal analysts across the political spectrum described this reading as far exceeding any authority Congress ever granted under these statutes.
What the Data Actually Shows About Anthropic's National Security Posture
The political characterization of Anthropic as a "woke company" run by "leftwing nutjobs" — Trump's words — does not survive contact with Anthropic's actual operational history. The data is worth presenting plainly, because it is not the story being told in most political coverage.
| Action | Direction | Cost to Anthropic |
|---|---|---|
| Cut off CCP-linked firms from Claude access | Pro-U.S. national security | Hundreds of millions in revenue |
| Shut down CCP-sponsored cyberattacks abusing Claude | Pro-U.S. national security | Operational cost + adversarial exposure |
| First AI lab deployed on classified Pentagon networks | Pro-U.S. national security | Significant engineering + compliance cost |
| Publicly advocated for strong chip export controls | Pro-U.S. technological advantage | Revenue from international customers |
| Refused autonomous weapons restriction removal | Contested — safety vs. capability | $200M+ contract + government access |
This is not the profile of a company that is hostile to American national security. It is the profile of a company that has repeatedly traded revenue for national security alignment — except on two specific use cases it considers categorically unsafe regardless of who is asking.
David Sacks, the White House AI and crypto czar, accused Anthropic of "regulatory capture based on fear-mongering." This is a coherent critique if one believes AI safety concerns are exaggerated. It is not coherent as a national security argument, because cutting off the CCP at hundreds of millions in cost is not the behavior of a company playing regulatory games for competitive advantage.
Why OpenAI Said Yes and Anthropic Said No
The Pentagon announced a deal with OpenAI the same weekend it designated Anthropic a national security threat. Sam Altman immediately posted on X inviting public questions about OpenAI's military work. The contrast was deliberate and pointed.
OpenAI accepted the "any lawful use" language. Anthropic did not. This is being framed as a story about values — Anthropic's founders care about safety, OpenAI has become more commercially pragmatic. That framing is partially true, but it misses a statistical point worth making.
The probability that AI causes catastrophic harm through military misuse is not zero. Anthropic's founding thesis — and the reason it commands a premium valuation over less safety-focused competitors — is that this probability is meaningful enough to warrant structural constraints. OpenAI's thesis has evolved: safety remains a stated priority, but operational flexibility for paying customers has been given increasing weight.
These are not just philosophical differences. They are different probability estimates about the same future, expressed as different contract terms. Anthropic is essentially saying: the expected cost of unrestricted military AI use, probability-weighted, exceeds the expected revenue from this contract. OpenAI, apparently, calculated the reverse.
"Both companies are making the same type of decision — a bet under uncertainty. They just disagree on the probabilities. That is a falsifiable disagreement. History will adjudicate it."
— The Alpha NodeThe Legal, Financial, and Strategic Outlook
Anthropic has stated unequivocally it will challenge the supply chain designation in court. The legal analysis from Lawfare — which tends to give significant deference to executive national security judgments — is unusually skeptical of the government's position. Three specific weaknesses stand out:
First: The designation exceeds statutory authority. FASCSA was designed for foreign adversary supply chain threats. Applying it to a domestic company over a contractual terms dispute is a novel use that will require the government to justify under heightened scrutiny.
Second: Hegseth's transition plan directly contradicts the urgency framing. Courts will ask: if Anthropic is an acute national security threat, why is it safe to keep using it in combat operations for six more months?
Third: Hegseth's public statements may have poisoned the litigation posture. Courts evaluating national security designations look for good-faith security rationale. Calling a company "woke" on X and describing its removal as a political act undermines the security justification the designation legally requires.
The financial damage to Anthropic, so far, appears more limited than the rhetoric suggests. Microsoft studied the designation and concluded Anthropic products remain available to its customers outside direct DoW contracts. The supply chain designation, as written under 10 U.S.C. § 3252, applies narrowly to Pentagon contract work — not to Anthropic's commercial business broadly. The nuclear interpretation — deplatforming Anthropic entirely by forcing AWS and Google Cloud to drop them — would require far more legal authority than exists.
The Deeper Question This Raises
This dispute is, at its core, about a question that no one in Washington has adequately answered: who sets the limits on military AI?
Anthropic's position is that private companies building transformative technology have a legitimate role in setting ethical use constraints — backed by contract, not just by government goodwill. The Pentagon's position is that the military's operational authority cannot be constrained by private company terms of service on matters of national defense.
Both positions have logic behind them. But the statistical argument for Anthropic's approach is underappreciated: when you are deploying a technology whose failure modes are not yet fully characterized, whose alignment properties under adversarial conditions are not fully understood, and whose misuse potential is asymmetric and potentially irreversible — the expected value of caution is higher than it looks in a single contract negotiation.
Autonomous weapons that misidentify targets do not generate refund requests. They generate casualties. The downside is not bounded. When the downside is unbounded, standard expected value calculations break down, and the rational response is to apply structural constraints regardless of the probability estimate on any individual deployment.
That is not a woke argument. That is a decision theory argument. And it is one the Department of War, so far, has not engaged.
The Bottom Line
Anthropic refused two contract terms. The government responded with the most punitive designation in its procurement history — a tool built for Chinese telecoms, applied to an American AI safety company. The legal case is weak. The financial damage appears limited. And Claude was used in combat operations the same weekend it was declared a national security threat. Draw your own conclusions about who is being consistent here.
Sources: CNBC (Mar 5–6, 2026) · Lawfare (Mar 1, 2026) · Axios (Feb 28, 2026) · TechCrunch (Mar 5, 2026) · Center for American Progress (Mar 5, 2026) · Verdict / Justia / Cornell Law Prof. Michael Dorf (Mar 3, 2026) · Anthropic official statements (Feb 27 – Mar 5, 2026) · ASIS Security Management (Mar 6, 2026)
Disclaimer: This is analytical commentary, not legal or investment advice. The Alpha Node applies statistical and logical frameworks to publicly available information.
한 줄 요약: Anthropic은 자율살상무기와 대국민 감시, 딱 두 가지를 거부했다 — 그리고 그 주말에 Claude는 이란 공습에 쓰이고 있었다.
Comments
Post a Comment