What did OpenAI agree with the DOD?

Permission to run in classified environments with safety strings attached

OpenAI reached an agreement with the Department of War to deploy its models inside a classified DOD network. The deal allows the agency to use OpenAI’s technology for some classified workflows while preserving constraints that OpenAI says are necessary to prevent use cases it considers dangerous or unethical.

According to company statements and reporting, the arrangement included several practical elements:

• a path for models to operate inside DOD’s classified systems rather than blanket external access;
• acceptance by the DOD of engineering and operational “red lines” proposed by the vendor, similar in spirit to limits other firms have tried to insist upon; and
• space for OpenAI to build and manage a bespoke safety stack rather than being forced to remove protections on demand.

The agreement was announced amid a broader industry showdown: Anthropic was locked in a public battle with the Pentagon over the same questions of guardrails and permitted uses. OpenAI’s deal reportedly differs because the DOD accepted contractual language that preserves some of the provider’s safety controls, and company leadership said the department would not compel the company to make the model comply if it judged a request unsafe. Sam Altman urged the Pentagon to apply similar terms industry‑wide.

Why it matters

This compact signals that the military can gain access to advanced models without forcing vendors to abandon their safety engineering. It creates a playbook for how classified deployments might be structured: private companies retain some control over model behavior, while the government gains the ability to run vetted systems inside secure networks. The outcome may influence which suppliers the Pentagon partners with, and whether other firms push to retain their own limits when negotiating national‑security contracts.