TLDR
- Pentagon officials want Anthropic to eliminate safety restrictions from Claude AI to enable unrestricted lawful applications, including autonomous weapons systems and mass surveillance operations.
- Dario Amodei, Anthropic’s CEO, declined the request, stating these applications could harm democratic principles.
- Pentagon officials established a Friday 5pm cutoff for Anthropic to accept the terms or lose access to defense contracts.
- Defense officials warned they might use the Defense Production Act to compel cooperation and designate Anthropic as a supply chain threat.
- Anthropic dismissed contract language delivered Wednesday evening as offering minimal meaningful changes.
Dario Amodei, CEO of Anthropic, has maintained his position on preserving safety restrictions for the company’s Claude AI system, placing a significant government partnership in jeopardy. Pentagon officials established a Friday cutoff time, requiring Anthropic to accept terms allowing “any lawful use” of the technology.
The disagreement focuses on two primary issues: deploying Claude for extensive domestic surveillance operations and operating completely autonomous weapon systems. According to Anthropic, these applications have always remained outside the scope of their Pentagon agreements and should remain excluded.
Amodei held discussions with Defense Secretary Pete Hegseth during the week. The talks concluded without resolution, prompting Pentagon officials to transmit updated contract terms Wednesday evening.
Anthropic dismissed the revised language. A company representative stated it demonstrated “virtually no progress” and contained legal terminology that could allow protections to “be disregarded at will.”
Pentagon officials have made their position clear through strong warnings. They indicated plans to remove Anthropic from defense partnerships and categorize the firm as a “supply chain risk” — a classification usually applied to vendors in adversarial countries.
A high-ranking Pentagon official also informed Reuters that Defense Secretary Hegseth might invoke the Defense Production Act. This legislation enables the government to compel companies to fulfill national defense requirements, regardless of their willingness. Legal scholars have raised doubts about whether such application of the statute would be legitimate.
What Anthropic Says About AI Weapons and Surveillance
In a published blog post, Amodei stated that contemporary AI systems, even the most sophisticated ones available, lack sufficient reliability to operate fully autonomous weapons. He emphasized that deploying these systems without human control creates dangers for military personnel and civilians alike.
Regarding surveillance capabilities, he cautioned that AI technology can “assemble scattered, individually innocuous data into a comprehensive picture of any person’s life — automatically and at massive scale.”
Anthropic expressed support for AI applications in lawful foreign intelligence operations while opposing domestic surveillance programs.
Pentagon representatives responded, with Undersecretary Emil Michael asserting that the applications concerning Anthropic already face prohibitions under existing laws and Pentagon regulations. Michael directly challenged Amodei on X, claiming he “wants nothing more than to try to personally control the US Military.”
The Business Risk for Anthropic
The commercial implications carry substantial weight. Pentagon officials have executed $200 million ceiling arrangements with prominent AI companies including Anthropic, OpenAI, and Google during the previous year.
Receiving a supply chain risk designation would prevent defense contractors like Lockheed Martin from incorporating Anthropic’s technology in Pentagon initiatives. The defense industrial ecosystem encompasses approximately 60,000 contractors.
Amodei proposed that Anthropic could collaborate with Pentagon officials on research and development efforts to enhance AI dependability for defense applications, though this proposal received no acceptance.
Through Thursday, both parties remained in a standoff with the 5:01 p.m. Friday cutoff time still active.
