Source: three AI Explained (Philip) YouTube auto-caption transcripts, fetched 2026-10-30. They are raw/Deadline_Day_for_Autonomous_AI_Weapons_Mass_Surveillance.md (youtube.com/watch?v=Cru804JMjPI, uploaded 2026-02-27, 13:39), the second half of raw/What_the_New_ChatGPT_5.4_Means_for_the_World.md (youtube.com/watch?v=zizoDORjmlQ, 2026-03-06, 21:51) and the opening of raw/Two_AI_Models_Set_to_stir_government_urgency_But_Will_This_Challenge_Undo_Them.md (youtube.com/watch?v=s4tptozUJ8Y, 2026-03-26, 16:27). Most claims are AI Explained relaying other outlets (Politico, The Verge, Bloomberg, Time, The Atlantic, The Washington Post, Axios). Each claim names its outlet. Creator: AI Explained Platform: YouTube

Anthropic faced a 5:00 p.m. deadline on Friday 27 February 2026. The US Department of War (the Pentagon) wanted Claude available for any “lawful” use. Anthropic refused two uses: fully autonomous weapons and mass domestic surveillance of Americans. Within a week it had been formally designated a supply-chain risk, OpenAI had taken the contract, and a leaked Amodei memo had turned it into a public fight between the labs. This article records the sequence as AI Explained reported it at the time. It is the early history behind the designation and litigation tracked in Anthropic’s Position, August 2026.

Key Takeaways

  • Anthropic’s two red lines: no Claude for fully autonomous lethal weapons without a human in the loop, and none for mass domestic surveillance of Americans. It said it had “never raised objections to particular military operations.”
  • Its reasons were practical, not only ethical. On surveillance, Anthropic conceded the practice may be legal, but “only because the law has not yet caught up” with AI’s ability to stitch scattered data into a full profile of a person. On weapons: “Frontier AI systems are simply not reliable enough to power fully autonomous weapons.”
  • The Pentagon made two threats that Anthropic called contradictory: designate Anthropic a supply-chain risk, a label never before used on an American company, and invoke the Defense Production Act to force removal of the safeguards. Anthropic’s question was how it could be an adversary and also essential to national security at the same time.
  • Outcome by 6 March: Anthropic was officially designated a supply-chain risk (“I fired them like dogs”, per the US president). OpenAI took the contract with what it called added safeguards. The Department of War gave itself six months to keep using Claude.
  • AI Explained’s own view: there are no clean heroes. Bloomberg reported that Anthropic had dropped its responsible-scaling guarantee two days before the deadline. The Washington Post reported that Claude inside a Palantir system had suggested and prioritised hundreds of targets in Iran.
  • By 26 March the deal was being revisited. Axios reported that the Pentagon was reconsidering, after Anthropic warned officials that its next model would “supercharge” offensive and defensive cyber capability.

Timeline

Up to 27 February: the deadline

  • The demand. The Department of War wanted “almost unfettered” use of Claude. The only limit was that use be “lawful”, which AI Explained argues still allows autonomous weapons and mass domestic surveillance.
  • Pentagon response. Under Secretary of War Emil Michael called Dario Amodei “a liar” with “a god complex” who “wants nothing more than to try to personally control the US military”, and said the Department “will always adhere to the law.”
  • Other labs. Per Politico, Google and OpenAI were “only close to agreeing” terms. xAI was complying.
  • Employee petition. An open letter at notdivided.org, “We will not be divided”, asked Sundar Pichai and Sam Altman to keep refusing the domestic-surveillance and autonomous-killing uses. It had about 340 Google and OpenAI signatures on the morning of 27 February and was rising during recording.

The five twists AI Explained lists

  1. An existing deal already said no. Anthropic already had an agreement under which the Pentagon accepted responsible-use terms: no AI-controlled autonomous weapons and no AI domestic surveillance of Americans. Claude was in wide use at the Pentagon, at defence contractors and at Palantir. The Verge framed the question as whether a company “can be compelled out of a policy that the government itself has committed to in principle.”
  2. Pentagon policy already restricts both uses. DoD Directive 3000.09 requires that autonomous weapon systems let commanders “exercise appropriate levels of human judgment over the use of force.” A separate directive and the Responsible AI implementation pathway bar intelligence components from collecting on US persons except under specific legal authorities.
  3. The threats contradicted each other. A supply-chain-risk designation from Pete Hegseth would bar contractors such as Palantir from using Claude on government work, costing Anthropic “hundreds of millions, if not billions”. A Defense Production Act order would force Anthropic to build an unrestricted Claude. Anthropic: “we cannot in good conscience accede to their request.”
  4. The objections were about reliability. See Key Takeaways. AI Explained cites two papers in support:
    • Agents of Chaos (about 84 pages): agents running in OpenClaw-style harnesses, tested on open-weight models (Kimi series) and Claude Opus, followed most instructions from people who were not their owners. They disclosed 124 email records. One agent refused a direct request for personal information but forwarded the email containing it, unredacted.
    • Towards a Science of AI Agent Reliability (Princeton): headline accuracy hides unreliability along four dimensions. Consistency is low variance across repeated trials. Robustness is holding up under small prompt or tool-call changes. Predictability is being able to anticipate outputs. Safety is about how severe the failures are. AI Explained summarises its finding that progress on these from o1 to GPT-5.2, and from Claude 3.5 to Opus 4.5, was “not exactly a dramatic one.”
  5. Anthropic had just softened its own commitment. Per Bloomberg, Anthropic had dropped two days earlier its responsible-scaling-policy pledge not to train a model unless safety measures were guaranteed adequate in advance. The pledge no longer applies if Anthropic believes it lacks a significant lead over a competitor. Co-founder Jared Kaplan told Time: “it wouldn’t actually help anyone for us to stop training AI models … if competitors are blazing ahead.” AI Explained points out that this is exactly the unilateral commitment Anthropic was making toward the Pentagon.

Other voices:

  • Dean Ball (caption: “Dean B”), lead author of the White House AI Action Plan, suggested the Pentagon was standing on principle and did not actually intend to cross the red lines.
  • A former DoJ–Pentagon liaison (caption: “Katie Sweten”, spelling unverified) disagreed: “if these are the lines in the sand … I would assume that one or both of those functions are scenarios that they would want to utilize.”

27 February – 6 March: designation, OpenAI’s contract, the memo

  • Designation. Anthropic was officially told it is a supply-chain risk. The US president: “I fired Anthropic … like dogs.”
  • OpenAI took the contract. Per Anthropic, talks failed because the Department of War wanted to keep the option of using Claude for what could become domestic surveillance and fully autonomous warfare. Sam Altman said OpenAI added safeguards against both and signed a contract worth hundreds of millions of dollars.
  • The Amodei memo (about 1,600 words, leaked):
    • Called OpenAI’s messaging “mendacious”.
    • Said the safety layer Anthropic was offered, a classifier supplied by Palantir that flags and denies some uses, was “maybe 20% real, but 80% safety theater” and “easily overridden”.
    • Quoted Palantir as saying the service’s purpose was to placate unhappy employees or make “what is happening invisible to them.”
    • Said Pentagon human-in-the-loop policy, unlike law, “can be changed unilaterally.”
    • Attributed the administration’s hostility partly to Anthropic not donating to Trump, noting that Greg Brockman had donated “around 25 million”.
    • Called OpenAI employees “a gullible bunch” because of selection effects. Some OpenAI staff then posted “feeling sort of gullible today”.
  • Amodei’s apology. About 24 hours before the 6 March video, he said the memo “probably shouldn’t have been published” and that its tone “does not reflect my careful or considered views.”
  • Cloud vs on-device. Per The Atlantic, the Pentagon offered that the models would stay in the cloud rather than inside weapons. OpenAI cited this as a safeguard. Anthropic’s view was that if the model makes battlefield decisions from AWS, it is “a distinction without a difference.”
  • Altman’s position. The military’s operational decisions are “up to the government”: “Maybe you think the Iran strike was good and the Venezuela invasion was bad … you don’t get to weigh in on that.” If OpenAI refused, at least one other actor would agree to anything.
  • The Pentagon closed talks. The Department of War said there would be no more negotiations. On the All-In podcast Emil Michael said Amodei wanted to be phoned about each exception to Claude’s guardrails.
  • Claude in Iran targeting. The Washington Post reported that Claude, inside a Palantir system, had suggested hundreds of targets in Iran, given precise coordinates and ranked them by importance. AI Explained notes this may not breach Anthropic’s terms, and that it fell inside the six-month window the Department of War allowed itself.
  • Google DeepMind was reportedly working with the Department of War but saying little about it.
  • Background. OpenAI changed its terms to allow military use in January 2024 and kept a ban on weapons development and harming people.

26 March: the deal is revisited

  • Cyber warning. Per Axios, Anthropic told US officials its next major model “will supercharge both offensive and defensive cyber capabilities”. That could push the government toward a deal, and the lead Pentagon negotiator said they were “still very close to an agreement.”
  • Speculation. Axios raised the idea that Anthropic might help fund Trump Accounts ($1,000 per enrolled newborn). Brad Gerstner, an adviser to Amodei, designed that programme.
  • Which model. The model behind the warning is most likely Claude Mythos Preview (2026-04-07), which Anthropic withheld from general release because of its cyber capability.

What this means for practitioners

  • Vendor terms can change with politics. Within one week, a lab went from a primary defence supplier to a designated supply-chain risk, and a rival took the contract. If your business depends on a single model provider, especially for government or regulated work, keep a tested fallback provider. Mythos federal shutdown is the later example of the same risk.
  • Ask vendors where their limits sit: model, contract or classifier. The dispute turned on the difference between restrictions in the contract and a “safety layer” classifier on top. Amodei called the classifier “80% safety theater”. When a vendor promises safeguards, ask which kind they are.
  • Measure reliability, not just accuracy, before giving agents real authority. The four Princeton dimensions (consistency, robustness, predictability, failure severity) make a usable checklist for any agent with access to email, money or production systems. In Agents of Chaos, agents obeyed people who were not their owners and leaked data through a forwarded email.
  • Policy is not law. Amodei’s point was that internal policy such as DoD Directive 3000.09 can be changed unilaterally. The same applies to a vendor’s acceptable-use policy and to a lab’s responsible-scaling commitments, as Anthropic’s own RSP change shows.

Open Questions

  • Exact wording of the RSP change. Reported through Bloomberg and Time. The policy text itself was not checked. Whether this is the change that produced RSP v3.0, under which Mythos Preview shipped, is unconfirmed.
  • Contract value and safeguards. “Hundreds of millions of dollars” is AI Explained’s figure. OpenAI’s actual contract terms were not reviewed.
  • Washington Post Iran-targeting report. Relayed second-hand. The underlying article was not read.
  • Caption-uncertain names. The former DoJ–Pentagon liaison (“Katie Sweten”) and the “Majuro raid”, which AI Explained describes as the operation to capture Venezuela’s leader and is presumably Maduro, are not verified.
  • Later status. Anthropic’s Position, August 2026 tracks the litigation. A ruling called the designation “arbitrary and capricious”, but sources disagree on which proceeding is live. This article supports that article’s previously unverified “March” contract-cut detail: by 6 March the designation was official.