Court Blocks Trump Administration's AI Sanctions Against Anthropic as Retaliatory and Unconstitutional

The Trump administration's approach to artificial intelligence regulation has drawn sharp criticism from civil liberties advocates for what they characterize as a contradictory and retaliatory strategy. While the administration has broadly minimized AI regulation in pursuit of maintaining global competitiveness in frontier model development, it has simultaneously pursued aggressive sanctions specifically targeting Anthropic, an AI safety-focused company.

The conflict escalated after Anthropic refused government demands to use its language models for autonomous weapons development and surveillance operations. In response, the Pentagon designated Anthropic a "supply chain risk," effectively banning federal agencies and contractors from using the company's services—a move that would have cost the company hundreds of millions of dollars. The administration further imposed export controls on Anthropic's Mythos and Fable models, forcing the company to shut down those systems entirely to comply with restrictions on foreign national access.

The Electronic Frontier Foundation and allied civil liberties organizations filed an amicus brief arguing that these sanctions violate the First Amendment, as they constitute punishment for the company's public refusal to comply with government demands rather than legitimate national security measures. A federal court issued a preliminary injunction blocking the sanctions from taking effect, accepting the civil liberties groups' constitutional arguments. The EFF characterized the government's actions as retaliatory and inconsistent with how other AI companies with similar capabilities are regulated.

Critics note a stark policy inconsistency: while the administration has rolled back safety regulations addressing serious threats such as AI-enabled cyberattacks on government systems, it has subjected only Anthropic to exceptional restrictions. Other large language models with comparable offensive cybersecurity capabilities remain subject only to voluntary, light-touch frameworks encouraging companies to submit models for testing 30 days before public release. The EFF and cybersecurity experts argue that the export controls on Mythos models lack factual grounding, as the purported dangers of finding and exploiting software vulnerabilities represent well-established capabilities, not unique threats warranting exceptional restrictions.