The OpenAI Department of War AI agreement has been formally detailed, outlining the contractual framework, deployment model, and safety architecture governing the company’s classified AI operations.

In a public explanation of the contract, OpenAI stated that the arrangement includes stronger guardrails than previous classified AI deployments. The company emphasized that it retains control over its safety systems and is deploying models exclusively within a controlled cloud environment.

At the center of the agreement are clearly defined red lines that shape how its models may be used in national security contexts.

OpenAI Military AI Safety Redlines

The OpenAI Department of War AI agreement establishes three core prohibitions.

First, OpenAI technology cannot be used for mass domestic surveillance. Second, it cannot independently direct autonomous weapons systems. Third, it cannot be used for high-stakes automated decision-making, such as systems resembling social credit frameworks.

According to the company, these OpenAI military AI safety redlines are shared in principle by several frontier AI labs. However, OpenAI argues that its layered enforcement mechanisms go further than relying solely on usage policies.

The agreement explicitly references existing U.S. laws and Department of Defense directives, including DoD Directive 3000.09 (January 25, 2023), which requires rigorous verification and human oversight for autonomous systems. It also references the Fourth Amendment, the National Security Act of 1947, the Foreign Intelligence Surveillance Act of 1978, Executive Order 12333, and the Posse Comitatus Act.

The contract language clarifies that AI systems may not independently direct autonomous weapons in cases where human control is required, nor may they assume high-stakes decisions that require human authorization.

OpenAI Cloud-Only Classified Deployment Safety Stack

A defining feature of the OpenAI Department of War AI agreement is its deployment architecture.

The company confirmed that this is a cloud-only classified deployment safety stack. Models will not be deployed on edge devices, eliminating the possibility of powering fully autonomous lethal systems outside centralized oversight.

OpenAI retains full discretion over its safety stack, which includes classifiers, monitoring systems, and ongoing updates. Cleared OpenAI engineers and safety researchers will remain in the loop, providing direct oversight and technical support within classified environments.

The company stated that this architecture allows independent verification that its red lines are not crossed. It also noted that it would not provide “guardrails off” or non-safety-trained models for defense use.

Why the Deal Was Reached

OpenAI explained that it initially declined to enter classified deployment agreements until safeguards were sufficiently mature. The company said it worked to ensure that any deployment would preserve its safety framework rather than weaken it for performance gains.

It also acknowledged a desire to de-escalate tensions between the Department of War and U.S. AI labs. As part of the arrangement, OpenAI requested that similar contractual terms be made available to other AI companies.

When asked why it was able to reach an agreement where others could not, OpenAI stated that its contract structure, cloud-only deployment, and enforceable redlines may have created clearer guarantees.

The company also stated that it does not believe Anthropic should be designated as a supply chain risk.

Enforcement and Contingencies

Under the OpenAI Department of War AI agreement, the company retains the right to terminate the contract if terms are violated.

The contract also references existing surveillance and autonomous weapons policies as they stand today. OpenAI stated that even if laws or policies change in the future, the use of its systems must remain aligned with the standards reflected in the agreement.

In its explanation, the company expressed confidence that, given the safety stack, cloud-only deployment model, and legal framework, its systems cannot be used to power autonomous weapons or conduct mass surveillance of U.S. persons.

OpenAI framed the agreement as an attempt to balance national security needs with democratic oversight, arguing that responsible collaboration between AI labs and government institutions is essential as advanced systems continue to evolve.