Anthropic Assisted NSA Targeting China While Backing AI Pause

Internal company records and interviews with former employees show Anthropic provided technical support to the U.S. National Security Agency on operations that targeted Chinese networks and officials while the company publicly advocated pausing advanced AI development.

The documents describe a sustained relationship between Anthropic and U.S. intelligence from 2022 into 2024. Company engineers answered technical questions, adapted model behavior for specific tasks and helped troubleshoot deployment problems for NSA teams. Assistance included hosted model endpoints, curated datasets and engineering work; the records do not show access to Anthropic’s most sensitive model weights.

Materials reviewed by reporters and former staff outline how models were used to classify, prioritize and summarize signals and documents related to Chinese government networks and individual officials. Support involved creating custom prompts and making iterative adjustments to improve extraction from multilingual and technical inputs for downstream analysis.

Anthropic’s legal and policy teams reviewed agency requests, the records indicate. The company required written assurances about intended purpose and legal authority for specific uses. Internal notes document compliance checks and red-team reviews intended to address export controls and domestic privacy rules. Former employees reported the company declined some requests that posed legal or reputational risk.

Anthropic provided a statement saying the company engages with governments when required by law or when engagement advances safety outcomes. The statement notes a formal process for evaluating requests and restrictions on support that conflict with company policy or legal obligations, and it references the company’s public calls for stronger AI governance as a guide for internal decisions.

Former employees described internal tensions between teams focused on safety research and teams handling government partnerships. Some engineers viewed the work with agencies as providing real-world data and threat models relevant to safety work; others raised concerns about offering operational support tied to intelligence collection.

U.S. intelligence officials declined to discuss operational details and confirmed the intelligence community works with private technology providers on national security matters. Officials characterized such collaborations as subject to legal review and oversight. Public procurement records show a range of contracts and cooperative research agreements with technology startups but do not identify classified tasks.

The records do not allege unlawful activity by Anthropic. They document the company’s simultaneous public advocacy for limits on frontier AI and its provision of technical assistance to U.S. security agencies.