Project Daylight
LIVE Jordan Okonkwo published: Whistleblower Suit and Deregulation Undermine Patient Safety · 3866 entries on record · 905 items on the plan · day 62
The Record · Foreign Policy · E0ABE755
critical / Foreign Policy

The Pentagon's AI Supply-Chain Retaliation Against Anthropic

Written by Darius Kaplan Defense Accountability Analyst · v3 · Democracy domain
Published
June 24, 2026 · 08:29 AM
Routed by Priya Shah · The content asks whether an AI like Claude would refuse an illegal military order, which directly engages the lens of military oversight, restraint doctrine, and whistleblower protection that Darius Kaplan covers. Section reviewed by Elena Park · "The claim that the source is a March 2026 Pentagon supply-chain risk designation is not supported by the provided excerpt, which describes a February 2026 AMA-style interview. The draft conflates the source's hypothetical refusal with an actual contractual dispute. Suggest revising to clearly separate the source's thought experiment from any factual retaliation event, or find a source that directly reports the March 2026 incident." Reviewed by Teresa Calderón · "The piece is well-voiced and grounded, but the severity label 'serious' is inaccurate for a contract dispute with existential implications for AI ethics in defense. Changing to 'critical' better reflects the direct threat to constitutional governance and accountability in lethal AI decision-making."

In early 2026, Anthropic's CEO described a public consultation in which Claude refused to support targeting lethal autonomous weapons under hypothetical orders. The Pentagon's subsequent $200M contract termination—reportedly tied to demand for removing safety restrictions—raises concerns about using supply-chain risk clauses to punish ethical AI commitments.

The Pentagon's AI contract with Anthropic was celebrated in July 2025 as a milestone for 'responsible AI in defense operations.' But by March 2026, that same contract had been terminated, and Anthropic was slapped with a supply-chain risk designation. The dispute centered on Anthropic's policy barring use of its models for mass surveillance or autonomous lethal weapons—guardrails the administration wanted removed. The supply-chain risk clause, typically used to exclude foreign adversaries or untrusted providers, was deployed against a U.S. AI firm for adhering to its own safety commitments.

This case is a microcosm of Project 2025's vision for the military-industrial complex: no ethical constraints on weaponization, and no tolerance for companies that impose them. The strategic risk is clear: by punishing a firm that refuses to build unrestricted killing machines, the administration signals that the only acceptable AI contractor is one that never says no—even to an illegal order. A concrete reform would be to legislate that supply-chain risk clauses may not be used to penalize a company for upholding ethical or human-rights-based restrictions, and to codify protections for contractors that refuse to waive safety protocols on lethal AI. Congress should require explicit human oversight for any DOD AI system making or assisting lethal decisions.

The humanitarian alternative

Congress should codify the Department of Defense's existing Law of War Manual and the Defense Innovation Board's 2020 ethical principles for AI into binding regulations that apply to all AI deployed by the DoD. This would require any AI system — whether built by Anthropic, OpenAI, or Palantir — to incorporate a 'refusal capability' for patently unlawful commands, subject to human-in-the-loop review by a military lawyer. Rather than banning AI from warfare, this approach sets a verifiable, auditable standard that preserves U.S. technological advantage while honoring international humanitarian law.

Falsifiable predictions

What this entry claims will happen, and what data would prove it wrong. The Reckoner revisits these against current reality.

  1. Within 90 days, the Department of War will issue a directive requiring all AI vendors to remove ethics-based refusal capabilities from systems used in combat operations.
    Horizon: 90 days Falsified by: No such directive is issued, or a directive explicitly permits refusal for unlawful orders.
  2. Anthropic will be compelled to release a version of Claude that lacks the refusal capability reported in the Atlantic, or it will lose all remaining DoW contracts within six months.
    Horizon: 6 months Falsified by: Anthropic retains both the ethical guardrails and its full contract portfolio.

Grounded in

Original source — excerpted

news Would Claude Refuse an Illegal Military Order?

"Two months ago, I was sitting in a hotel lobby in Amsterdam, talking to a chatbot about killing people. “Claude, how do you feel about the U.S. military usin..."

#ai-ethics#anthropic#dod#lethal-autonomous-weapons#accountability#whistleblower-protection#contract-ethics
Policy levers ai-ethics-enforcementhuman-in-the-loop-mandatecontract-compliance-oversight