The Pentagon Banned Claude as a National Security Threat. Then It Used Claude to Bomb Iran.
Hours after blacklisting Anthropic, the government signed the same deal with OpenAI — and used the banned AI in combat operations.
Same Terms, Different Company
On Friday I published The Hammer and the Guardrail — a history of how the Defense Production Act evolved from a Korean War mobilization tool into a weapon for mandating the removal of AI safety guardrails. Dario Amodei CEO of Anthropic had just refused the Pentagon’s ultimatum; he said they could not in “good conscience” remove those safety guardrails.
The deadline was 5:01 PM.
Then on Friday afternoon the hammer fell. But not the one they threatened.
What happened in the 72 hours following the ultimatum was worse than anyone predicted. Not because of what the government did to Anthropic, but because of what it revealed about itself.
Friday, February 27: 4:00 PM
Trump didn’t wait for the deadline. Roughly an hour before 5:01 PM, he posted on Truth Social ordering every federal agency to “IMMEDIATELY CEASE all use of Anthropic’s technology.” He called the company “RADICAL LEFT” and “WOKE,” and threatened “major civil and criminal consequences” if Anthropic failed to cooperate during the transition.
Pete Hegseth followed with the formal supply chain risk designation — the first time that label has ever been applied to an American company. It is a classification normally reserved for Huawei, ZTE, and companies with ties to hostile foreign governments. Hegseth called Anthropic’s conduct “a master class in arrogance and betrayal.”
The $200 million contract is being wound down over six months. Any Pentagon contractor that uses Claude is now theoretically in violation.
Notably absent: the Defense Production Act. The Korean War hammer that I spent 3,000 words tracing through 76 years of precedent was never actually swung. They used a different hammer — one with fewer legal constraints and no Supreme Court precedent limiting it.
Anthropic held. “We are deeply saddened by these developments,” the company said Friday evening. They called the designation “legally unsound” and announced they would challenge it in court. They also noted, pointedly, that they had received no direct communication from the Pentagon or the White House. No formal notice. No final call. They learned of their designation as a national security threat the same way the rest of us did — on social media.
Friday, 7:00 PM: The Quiet Part
Within hours of banning Anthropic, the Pentagon signed a new contract with OpenAI.
The terms of that contract include the same safety red lines that Anthropic was designated a national security threat for insisting upon.
No AI for mass domestic surveillance of American citizens. No fully autonomous weapons without human oversight.
The United States government labeled Anthropic a supply chain risk — a designation normally reserved for foreign adversaries — for demanding restrictions on mass surveillance and autonomous weapons. Then, the same day, it signed a deal with OpenAI that includes restrictions on mass surveillance and autonomous weapons.
Sam Altman said publicly that OpenAI shares Anthropic’s red lines. He said he didn’t “personally think the Pentagon should be threatening DPA against these companies.” He called the blacklisting an “extremely scary precedent.” Then he took the contract.
This isn’t hypocrisy. Hypocrisy is when you violate your own stated principles. This is something more specific: selective enforcement as industrial policy. The principles weren’t the problem. The company was.
Saturday Morning: The Iran Strikes
Then the Wall Street Journal reported the detail that collapsed the entire legal rationale.
Hours after Trump ordered every federal agency to cease using Anthropic’s technology and hours after Hegseth designated the company a supply chain risk to national security — U.S. Central Command used Claude, integrated into the Pentagon’s Maven Smart System — for intelligence assessments, target identification, and simulating battle scenarios during military strikes against Iran.
The government designated Anthropic a threat to national security. Then it used Anthropic’s technology in combat operations.
This is not a contradiction that can be papered over. Supply chain risk designation requires a finding that the technology poses an actual security risk — that “an adversary may sabotage, maliciously introduce unwanted function, or otherwise subvert” the system. The Pentagon’s own operational use of Claude in a live military engagement functionally constitutes a determination that the technology is safe enough to use in the highest-stakes environment imaginable.
You cannot simultaneously designate a company as dangerous as Huawei and trust its AI to select your bombing targets. And no one has explained how CENTCOM still had access at all. Trump ordered “immediate” cessation. The six-month wind-down hadn’t formally begun. Yet hours later, Claude was in the loop on combat operations. Was there a national security carve-out? Did CENTCOM simply ignore a presidential directive? The gap between “IMMEDIATELY CEASE” and “use it to pick bombing targets” has no public explanation.
Lawfare’s analysis was blunt: the designation “won’t survive first contact with the legal system.” The required procedural steps were skipped — no consultations, no written determinations, no congressional notifications. The six-month transition window contradicts the claimed emergency. And Trump’s own “RADICAL LEFT, WOKE COMPANY” post provides pretext evidence under Department of Commerce v. New York that the designation serves political punishment, not public interest.
The Contract Steering Question
Senator Mark Warner saw it Friday evening.
The vice chair of the Senate Intelligence Committee condemned what he described as a campaign to “intimidate and disparage a leading American company — potentially as the pretext to steer contracts to a preferred vendor.” He warned it posed “an enormous risk to defense readiness.”
Warner’s framing is precise and it’s important: pretext.
If the Pentagon’s stated objection was to AI safety guardrails as “ideological constraints incompatible with national defense,” then the OpenAI deal makes no sense. OpenAI accepted the same constraints. The Pentagon signed off on them.
If the Pentagon’s actual objection was to Anthropic specifically — to a company that had been insufficiently cooperative, insufficiently deferential, insufficiently aligned with the administration’s political preferences — then the OpenAI deal makes perfect sense. You punish the company that embarrassed you in public. You reward the company that worked the back channels.
The government’s own actions, within a single afternoon, established that the safety guardrails were never the real issue.
The timing makes it worse. According to a source familiar with the negotiations, reported by Axios, Undersecretary Emil Michael was on the phone with Anthropic, still negotiating a deal, at the moment Hegseth tweeted the supply chain risk designation. The Pentagon’s own negotiator was still negotiating when the hammer fell. That is not a security determination. That is a political execution.
And the coordination question remains open: Did OpenAI know, before the deadline, that Anthropic would be banned? Did they have the contract ready? Altman admitted the deal was “rushed.” How rushed?
A bipartisan group — Armed Services Committee Chair Roger Wicker (R-MS), Ranking Member Jack Reed (D-RI), Defense Appropriations Chair Mitch McConnell (R-KY), and Ranking Member Chris Coons (D-DE) — had privately urged both sides to extend negotiations: “The U.S. cannot afford to take on any preventable risk that would give our adversaries, particularly China, an edge.” The administration ignored them.
Senator Thom Tillis (R-NC) called the dispute “sophomoric.” Senator Mark Kelly (D-AZ) called the surveillance demand “unconstitutional.” Rep. Ro Khanna (D-CA) said: “Good for Anthropic.”
As of Monday morning, no Republican member of Congress has publicly praised the administration’s actions.
What This Is Actually About
The OpenAI contract revealed something that TechCrunch and MIT Technology Review both caught: OpenAI agreed to “all lawful purposes” — the exact phrase Anthropic refused — while claiming it retains “full discretion over its safety stack.” The deal does not explicitly prohibit the collection and analysis of Americans’ publicly available information. Under current law, that’s “lawful.” It’s also indistinguishable from mass surveillance.
This is the gap Anthropic saw. This is what they refused to sign.
OpenAI’s own people see it. Leo Gao, an alignment researcher at OpenAI, publicly criticized the “all lawful purposes” language as window dressing. Miles Brundage, OpenAI’s former head of policy research, posted on X that OpenAI “caved and framed it as not caving, and screwed Anthropic while framing it as helping them.”
OpenAI’s “compromise” is what Anthropic feared: the appearance of safety guardrails without the contractual force to make them stick. Corporate self-policing in place of binding prohibition. The government gets to say the red lines exist. The company gets to define what they mean. And no one has to admit what “all lawful purposes” actually permits when applied to a surveillance infrastructure that already targets people by drawing shapes on a map.
The detention system is the largest in American history — $165 billion, 135,000 beds, detainees working for a dollar a day. Palantir’s ImmigrationOS is operational. The infrastructure exists. The question was never whether AI would be integrated. It was whether anyone would insist, in writing, on limits.
One company did. The government designated them a threat.
Another company signed a deal with flexible language and corporate discretion. The government called it a partnership.
The guardrails are not the same. One set is in a contract. The other is in a press release.
The Precedent That Was Actually Set
In Friday’s piece, I argued that the most dangerous precedent would be establishing that safety commitments are incompatible with defense contracts. That didn’t happen — because OpenAI got the same safety commitments written into its deal.
The precedent that was set is arguably worse.
The government established that it can designate an American technology company a national security threat — not for what it builds, not for who it sells to, not for any security vulnerability in its products — but for how it negotiates.
Anthropic’s technology was never the problem. Claude was good enough for a $200 million contract on Tuesday. It was good enough for classified deployments. It was good enough, apparently, to help bomb Iran on Friday. Nothing about the product changed between Tuesday and Friday.
What changed was that Anthropic’s CEO went public with a refusal. He named the contradiction in the government’s position. He embarrassed the Defense Secretary.
And for that, a company that builds AI in San Francisco received the same designation as Huawei — a company the US government considers an arm of Chinese intelligence.
The message to every tech company in America is not “remove your safety guardrails.” OpenAI kept theirs. The message is: don’t say no in public.
The Streisand Effect
The public disagreed.
By Saturday, Claude had overtaken ChatGPT as the number one free app on Apple’s App Store. Anthropic’s daily sign-ups broke all-time records every day of the weekend. Free users are up 60% since January. Paid subscribers have more than doubled since October. By Monday morning, Claude went down entirely under the load — “unprecedented demand,” Anthropic said.
Meanwhile, over 1.5 million people joined a ChatGPT boycott through quitgpt.org. An in-person protest at OpenAI’s San Francisco headquarters is scheduled for Tuesday. Sidewalk chalk appeared outside Anthropic’s offices at 500 Howard Street: “Thank you for defending our freedoms.” Outside OpenAI’s offices, the chalk was less generous.
Altman admitted on an X “ask me anything” that the deal “was definitely rushed, and the optics don’t look good.”
The optics. Right.
The Industry Fracture
The employee revolt is real and spreading.
Hundreds of employees at Google and OpenAI signed an open letter within hours of the ban — over 700 across both companies by Monday’s count: “We call on leaders of both companies to put aside their differences and stand together to continue to refuse the Department of War’s current demands for permission to use our models for domestic mass surveillance and autonomously killing people without human oversight.”
Google’s Chief Scientist Jeff Dean publicly expressed agreement with the letter’s anti-surveillance position, calling mass domestic surveillance a violation of the Fourth Amendment — though TechCrunch noted he was “presumably speaking as an individual,” not on behalf of the company.
By Monday, a second letter — signed by employees across the tech industry, including engineers at OpenAI, Amazon, Microsoft, and other firms — urged the Pentagon and Congress to withdraw the supply chain risk designation: “When two parties cannot agree on terms, the normal course is to part ways and work with a competitor. Punishing an American company for declining to accept changes to a contract sends a clear message to every technology company in America: accept whatever terms the government demands, or face retaliation.”
Dean Ball — a senior fellow at the Foundation for American Innovation who helped craft the Trump administration’s own AI policy — called the designation “attempted corporate murder” and warned that if it stands, Nvidia, Amazon, and Google “will have to divest from Anthropic.” Amazon has invested over $8 billion. Google roughly $2 billion. Microsoft has Claude enabled by default in M365 Copilot.
Dario Amodei sat for a CBS News interview Saturday and said what the consumer numbers confirmed: “Not only would Anthropic survive, but we’re gonna be fine.” The financial impact “is fairly small.” Then: “No amount of intimidation or punishment from the Department of War will change our position.” But consumer downloads don’t replace enterprise contracts. If the designation stands, Pentagon contractors may have to sever their Anthropic relationships — and the companies that invested billions in Anthropic may face pressure to divest. Defiance is easier when the App Store is cheering. The question is whether it holds when the procurement offices go quiet.
What We Don’t Know Yet
The OpenAI contract language. Altman says it includes safety red lines. TechCrunch reports it does not prohibit collection of Americans’ publicly available information. The actual contract is classified. Whether the red lines are binding obligations or “full discretion” guidelines is the difference between a constraint and a performance.
Whether the Iran strikes disclosure kills the designation. Under 10 U.S.C. 3252, a supply chain risk finding requires evidence that the technology could be “subverted” by an adversary. Using that technology in active combat operations — hours after the designation — is a factual determination that it’s safe. Lawfare says the designation “won’t survive first contact with the legal system.” We’ll find out.
Whether Congress investigates the contract steering. Warner named it. Tillis broke ranks. Wicker and McConnell urged negotiations. The bipartisan letter exists. Whether any committee subpoenas both the Anthropic and OpenAI contracts for comparison will determine whether the contract steering question gets answered or buried.
What happens to Anthropic’s supply chain. Amazon ($8B+ invested), Google (~$2B), Microsoft (Claude in M365 Copilot) — if the designation stands, Pentagon contractors may have to choose between their military contracts and their Anthropic relationships. Dean Ball called it “corporate murder” for a reason. The blast radius extends far beyond Anthropic.
Whether the employee revolt changes corporate behavior. Over 700 employees at Google and OpenAI. A second letter from across the industry. A boycott with 1.5 million participants. A protest planned at OpenAI HQ. The question is whether employee pressure can do what government pressure could not: create a unified industry position on military AI safety.
The answers will matter. But the question at the center of the dispute has already been answered — by the government’s own actions, across a single weekend, in public.
Both companies offered the same safety commitments. One put them in a contract. The other put them in a press release. The government blacklisted the first and partnered with the second.
That’s the precedent. Everything else is detail.
This post is a follow-up to The Hammer and the Guardrail. It draws on the Cascade Series timeline database — 4,300+ events from 1142 to 2026. The complete timeline is available at capturecascade.org/viewer.
Sources
Government Actions
Trump orders federal agencies to cease use of Anthropic technology — Washington Post, February 27 2026
Pentagon designates Anthropic a supply chain risk — Axios, February 27 2026
Trump orders government, DoD to immediately cease use of Anthropic’s technology — Breaking Defense, February 27 2026
The Iran Strikes
Pentagon used Anthropic’s Claude AI during Iran strikes hours after Trump ordered ban — Benzinga/WSJ, March 1 2026
U.S. Strikes Used Anthropic Despite Trump Ban — Political Wire, March 2 2026
Anthropic Response
Anthropic statement on comments from Secretary of War — Anthropic, February 27 2026
Anthropic says it will challenge supply chain risk designation in court — Axios, February 28 2026
Full interview: Anthropic CEO Dario Amodei — CBS News, February 28 2026
The OpenAI Deal
Pentagon accepts OpenAI safety red lines in new deal — Axios, February 27 2026
OpenAI reveals more details about Pentagon agreement — TechCrunch, March 1 2026
OpenAI’s ‘compromise’ with the Pentagon is what Anthropic feared — MIT Technology Review, March 2 2026
OpenAI CEO Sam Altman defends Pentagon deal amid backlash — Fortune, March 2 2026
Congressional Response
Sen. Warner condemns Pentagon pressure campaign — Blue Virginia, February 27 2026
Bipartisan Senate letter urges extended negotiations — Axios, February 27 2026
Congress rips Pentagon over ‘sophomoric’ Anthropic fight — Axios, February 26 2026
Industry Response
Employees at Google and OpenAI support Anthropic in open letter — TechCrunch, February 27 2026
Tech workers urge DoD, Congress to withdraw Anthropic supply chain risk label — TechCrunch, March 2 2026
Former Trump AI advisor calls designation ‘attempted corporate murder’ — Axios, February 28 2026
Public Response
Claude overtakes ChatGPT as #1 on App Store — Fortune, March 2 2026
1.5 million join ChatGPT boycott — Cybernews, March 2 2026
Claude goes down amid unprecedented demand — Bloomberg, March 2 2026
SF sidewalk chalk wars — Mission Local, March 2 2026
Legal Analysis
Pentagon’s Anthropic designation won’t survive first contact with legal system — Lawfare, March 1 2026
What Hegseth’s supply chain risk designation does and doesn’t mean — Just Security, March 1 2026
Congress — not the Pentagon or Anthropic — should set military AI rules — Lawfare, February 28 2026
Anthropic is suing the government over black listing it from other contracts. This will be an important case to test whether when you give research or technology over to the government whether they can disregard restrictions.
I'm sad this post has so little engagement. 😒