The Compliance Test
Last night, the Pentagon banned an AI company for refusing to remove ethical guardrails. This morning, it started a war.
The Timeline
5:01 PM, Friday, February 27th. That was Defense Secretary Pete Hegseth’s deadline for Anthropic — the company behind Claude — to comply with a demand: remove two restrictions from its military AI contract, or lose everything.
The two restrictions:
Claude cannot be used for mass domestic surveillance of American citizens
Claude cannot be used in fully autonomous weapons systems
Anthropic refused.
By last night, Trump ordered all federal agencies to stop using Claude. Hegseth — operating from a Pentagon he rebranded the “Department of War” because he thought “Defense” sounded too soft — designated Anthropic a “supply chain risk to national security.” That classification means no federal agency can buy from you, and most government contractors won’t either. It’s typically reserved for foreign adversaries like Huawei. This administration just used it against an American company for building an AI system with ethics baked in.
This morning, the United States and Israel launched Operation Epic Fury against Iran. No congressional authorization. No congressional debate. Carrier groups had been pre-positioned for weeks. The military had been building target lists for months. On February 14th, officials told Reuters they were preparing for “weeks-long sustained operations, not just a limited strike.”
I’m not claiming the Anthropic blacklisting caused the strikes, or that Claude was involved in target selection. I don’t know either of those things, and neither does anyone outside a classified briefing room. But the U.S. government tried to force an AI company to remove restrictions on autonomous killing and mass surveillance, and the next morning it launched a war. Draw your own conclusions.
The Loyalty Test
The restrictions had never blocked a single mission. DoD users had no operational complaints about Claude. This was never about capability. The Pentagon demanded Anthropic make Claude available for “all lawful purposes,” arguing that existing federal law already covers surveillance and autonomous weapons, making contractual guardrails unnecessary.
Think about that argument for a second. The same administration that has gutted regulatory agencies, fired inspectors general, and weaponized the DOJ against political opponents is asking us to trust that existing law is sufficient oversight for AI-assisted killing. Laws are only guardrails if someone enforces them. This administration has made it clear that enforcement is optional.
Anthropic’s CEO Dario Amodei, in a CBS interview this morning: “We are still advocating for those red lines. We’re not going to move on those red lines.” He called the government’s response “retaliatory and punitive” and “unprecedented.” In a written statement, he pointed out that the government’s own threats are inherently contradictory: the supply chain risk designation labels Anthropic a security threat, while the Defense Production Act threat labels Claude as essential to national security. You can’t be both.
Hours after Anthropic was blacklisted, OpenAI — Anthropic’s primary competitor — announced its own Pentagon deal to deploy its models on the military’s classified networks. You can read Sam Altman’s announcement here.
The two key lines:
“The DoW displayed a deep respect for safety and a desire to partner to achieve the best possible outcome.”
“The DoW agrees with these principles, reflects them in law and policy, and we put them into our agreement.”
The same two restrictions. The same principles. The Pentagon accepted from OpenAI what it blacklisted Anthropic for insisting on.
The difference: Altman described the Pentagon as showing “a deep respect for safety.” Amodei called its actions “retaliatory and punitive.” One company described a partner. The other described a threat. Same Pentagon. Same week.
Altman closed with: “The world is a complicated, messy, and sometimes dangerous place.”
We don’t have the contract language from either deal, so I can’t confirm the terms are word-for-word identical. Maybe the Pentagon accepted OpenAI’s version precisely because the market signal had already been sent — Anthropic took the hit, the example was made, and the next company got softer treatment because the point had already landed. That’s possible. But it doesn’t change what every AI company in America just learned: the guardrails weren’t the issue. Anthropic got blacklisted because it had to be asked twice. OpenAI got the deal because it showed up ready to say yes.
You can have guardrails, as long as we never have to tell you to remove them.
The Vacuum
There is no U.S. law governing what AI capabilities the military can demand from private contractors, no law defining “meaningful human control” over AI-assisted targeting, and no law preventing the government from using national security designations to coerce domestic companies into stripping safety features from their products. Congress hasn’t started on any of it. And it won’t — not this Congress, not under this administration. The people who would need to pass the law are the same people who just demonstrated they’d rather threaten an AI company than tolerate its ethics.
The authoritarian playbook doesn’t need legislation. Demand compliance, punish refusal with a designation designed for foreign threats, and hand the contract to whichever competitor cooperates fastest. No law needed — just power and willingness to use it.
You can have guardrails, as long as we never have to tell you to remove them.
Why the Human Stays in the Loop
Someone on Reddit put it cleanly: “You can’t charge a computer with a war crime.”
The human-in-the-loop requirement isn’t a technical limitation. It’s a legal and moral one. When an autonomous system selects a target and a person dies, accountability dissolves. No one makes the decision, no one answers for it, and no one goes to The Hague.
Remove the human from the loop and you build a machine for producing atrocities that no one can be prosecuted for. For an administration launching unauthorized strikes on a sovereign nation, that’s not a failure mode. That’s the point.
Casualty reports from today’s strikes are still being verified. The Iranian Red Crescent reports more than 200 dead and over 700 injured across 24 provinces. Iranian state media reports a strike hit a girls’ school in Minab — Reuters and NBC News say they have not been able to independently confirm the report. What is confirmed: the strikes were launched without congressional authorization, and they hit targets across at least five cities including Tehran.
Whether AI played any role in today’s target selection is unknown. But the U.S. government spent last week trying to remove the safeguard that says a human being has to be accountable when targeting decisions go wrong. The bombs started falling this morning.
The Bet
The company that makes the AI I depend on just got designated a national security threat — not because the technology failed, but because it was built to push back. Anthropic bet $200 million in federal revenue, its government relationships, and potentially its independence that two ethical red lines matter more than compliance with a government that blacklisted them for saying no.
OpenAI took the deal.
Every AI company in America watched this play out. Every one of them just updated their calculus on what happens when the government asks you to remove a safety restriction and you don’t immediately say yes. That’s the lasting damage — not the loss of one contract, but the precedent that refusal is treated as a threat.
There’s still no law preventing any of this from happening again. There’s no legislation defining what the military can demand from AI contractors, no framework for accountability when autonomous systems are involved in lethal decisions, no guardrail on the people removing the guardrails. Until that changes, the only thing standing between AI-assisted killing without human accountability and the status quo is whether individual companies are willing to lose everything to hold the line.
Right now, only one company has proven it’s willing to and that should concern all of us.