OpenAI announced a new agreement with the Department of War that adds explicit language banning domestic surveillance of U.S. persons. The update was posted on March 2, 2026 and clarified that the Department cannot use OpenAI tools for tracking or monitoring Americans, even if it acquires commercial data OpenAI blog.
The added clause reads: "The AI system shall not be intentionally used for domestic surveillance of U.S. persons and nationals," and repeats that "deliberate tracking, surveillance, or monitoring … is prohibited." OpenAI says the Department of War also agreed that its intelligence agencies, including the NSA, would need a fresh contract before accessing the models. The language mirrors the Fourth Amendment, the National Security Act of 1947, and the FISA Act of 1978 OpenAI blog.
OpenAI is not just putting a line in a contract. The company says it will keep full control over its safety stack, run the deployment in a cloud‑only environment, and keep cleared OpenAI engineers in the loop. That architecture lets the lab verify that the red lines are not crossed, including by running classifiers that flag prohibited use OpenAI blog.
Three red lines guide the deal:
1. No mass domestic surveillance. 2. No autonomous weapon direction. 3. No high‑stakes automated decisions like "social credit" systems.
OpenAI claims other frontier labs have weakened similar safeguards and rely mainly on usage policies. By contrast, OpenAI says its multi‑layered approach adds contractual protections, discretionary safety controls, and human oversight OpenAI blog.
The contract also references DoD Directive 3000.09, which requires rigorous verification, validation, and testing for any AI used in autonomous or semi‑autonomous systems. For intelligence work, the agreement ties compliance to the Fourth Amendment, the Foreign Intelligence Surveillance Act, Executive Order 12333, and the Posse Comitatus Act OpenAI blog. In short, the Department can use the models for lawful purposes, but it cannot run them unsupervised on lethal weapons or on decisions that demand a human sign‑off.
Why does this matter?
The Pentagon has been courting AI labs for classified deployments, and OpenAI's stance could set a benchmark. If the safeguards hold, the military gets access to state‑of‑the‑art language models while civilian privacy remains protected. If the clauses are ignored, the precedent could erode the privacy shield for all future AI contracts.
I'm skeptical about enforcement. The agreement gives OpenAI full discretion over its safety stack, but the Department will still have operational control over the cloud environment. Verifying compliance will likely depend on OpenAI's internal monitoring tools, which the public cannot audit. The working group the Department plans to convene, mixing lab leaders, cloud providers, and policy staff, might improve transparency, but it also creates a closed circle where oversight is limited to insiders.
The deal is more detailed than Anthropic's recent Pentagon contract, which reportedly left many safety provisions to policy language alone. OpenAI's claim that its guardrails are "more expansive" seems plausible, but the real test will be the first classified deployment. Until we see logs or third‑party audits, the promise remains just that, a promise.
Nobody knows how this will play out in practice. The agreement is a step toward formalizing AI use in national security, yet the gap between contract language and on‑the‑ground enforcement could be wide. We'll have to watch for any reports of misuse and for how the working group addresses emerging risks.