Anthropic Refuses To Drop AI Safeguards in Dispute With U.S. Department of War

What Anthropic Announced

On February 26, 2026, Anthropic CEO Dario Amodei published a public statement describing an active policy dispute with the U.S. Department of War. Anthropic said it wants to continue supporting U.S. national security operations and described Claude as already deployed for intelligence analysis, modeling and simulation, operational planning, and cyber operations. At the same time, the company said it will not remove two specific safeguards from its contracts.

Anthropic framed the position as consistent with a broader strategy: support democratic governments, provide high-capability AI to defense and intelligence users, and maintain explicit limits for high-risk use cases that the company believes are incompatible with civil liberties or current model reliability.

The Two Guardrails Anthropic Says It Will Keep

• Mass domestic surveillance: Anthropic said large-scale domestic monitoring enabled by frontier AI is incompatible with democratic values.
• Fully autonomous weapons: Anthropic said current frontier systems are not reliable enough to safely automate target selection and engagement without meaningful human control.

The company stated that these limits have long been part of its Department of War contracting terms and said it has offered R&D collaboration to improve reliability for future systems.

Why This Matters for AI Procurement

The statement indicates that procurement negotiations are moving beyond model performance into governance terms, liability boundaries, and human oversight requirements. Anthropic also said the Department threatened stronger escalation options if safeguards were not removed, while Anthropic argued those pressures do not change its position.

For the broader AI market, this is a significant test case. Frontier model vendors are increasingly expected to support defense workloads, but governments and providers may not align on what safeguards are mandatory versus optional. Anthropic's stance suggests that contract-level policy controls are becoming a primary mechanism for handling sensitive deployment risks in real-world public sector AI programs.

The near-term question is whether agencies adapt procurement frameworks to accommodate providers that keep explicit usage restrictions, or whether contracts shift toward any lawful use language as a baseline requirement. Either way, this dispute marks a new phase in how AI capability, national security demand, and governance constraints interact in production deployments.