OpenAI's Pentagon Deal Criticized as 'Sloppy'
OpenAI's expanding work with the U.S. military is drawing scrutiny, with CEO Sam Altman himself admitting the initial Pentagon deal “looked opportunistic and sloppy.” A new report details how OpenAI caved to pressure on AI surveillance, and the company just secured another deal to deploy models on classified networks for the Department of Defense.
The deal materialized just hours after the Pentagon terminated its relationship with rival AI lab Anthropic. The government designated Anthropic a "supply chain risk" and ordered federal agencies to cease using its technology after the company refused to remove contractual "red lines" prohibiting the use of its AI for mass domestic surveillance or for autonomous weapons systems that operate without human oversight. Anthropic's now-canceled contract was worth up to $200 million and had made its model, Claude, the only frontier AI model integrated into the military's most sensitive networks. The company’s refusal to accept the Pentagon's desired "all lawful purposes" standard for a renewed contract led to the breakdown. In contrast, OpenAI agreed to this standard before amending its contract after public outcry. The backlash against OpenAI was swift, leading to a "delete ChatGPT" campaign online and a surge in downloads for Anthropic's Claude, which briefly topped Apple's App Store charts. The controversy also prompted an open letter from over 430 employees at major tech companies, including Google and Microsoft, urging their employers not to agree to the Pentagon's demands. One OpenAI alignment researcher characterized the initial safeguards in the deal as "not really operative except as window dressing." In response to the criticism, OpenAI amended its agreement to explicitly forbid its models from being used for intentional domestic surveillance, including the analysis of commercially acquired personal data. The updated terms also prevent intelligence agencies like the National Security Agency (NSA) from using the technology without a new, separate contract. OpenAI's approach to enforcing these "red lines" differs from Anthropic's. While Anthropic insisted on explicit prohibitions written into the legal contract, OpenAI's solution is more technical, relying on a cloud-only deployment, retaining control over its internal "safety stack," and having cleared OpenAI personnel involved in oversight. The company's policy on military engagement has evolved. In January 2024, OpenAI quietly removed a broad ban on "military and warfare" applications from its usage policy, replacing it with a more general prohibition against using its services to cause harm. This change paved the way for deeper collaboration with defense agencies.
