---
type: "post"
id: "openai-says-it-reached-an-agreement-with-the-b5a23c"
slug: "openai-says-it-reached-an-agreement-with-the-b5a23c"
title: "OpenAI says it reached an agreement with the Department of War to deploy advanced AI systems in classified"
description: "OpenAI says it reached an agreement with the Department of War to deploy advanced AI systems in classified environments, with a cloud-only architecture, retained safety controls and human oversight."
retrieval_nugget: "OpenAI says it reached an agreement with the Department of War to deploy advanced AI systems in classified environments, with a cloud-only architecture, retained safety controls and human oversight."
published_at: "2026-11-18"
updated_at: "2026-11-18"
record_date: "2026-11-18"
date_kind: "published_at"
topics: ["anthropic"]
entities: ["OpenAI","Department of War","Anthropic","Daybreak Red Team","National Security Agency"]
related_patterns: ["agent-security-moves-to-runtime-boundaries"]
novelty_label: "new_fact"
what_changed: "OpenAI says it added explicit limits barring domestic surveillance of U.S. persons, including through commercially acquired data, and excluded intelligence agencies such as the NSA absent a new agreement."
prior_refs: ["https://newruntime.com/posts/strengthening-nato-and-allied-cyber-defenses-with-ai-b3a3de/"]
source_url: "https://openai.com/index/our-agreement-with-the-department-of-war"
source_urls: ["https://openai.com/index/our-agreement-with-the-department-of-war"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:openai-says-it-reached-an-agreement-with-the-b5a23c"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/openai-says-it-reached-an-agreement-with-the-b5a23c/","markdown":"https://newruntime.com/posts/openai-says-it-reached-an-agreement-with-the-b5a23c.md","json":"https://newruntime.com/posts/openai-says-it-reached-an-agreement-with-the-b5a23c.json"}
markdown_url: "https://newruntime.com/posts/openai-says-it-reached-an-agreement-with-the-b5a23c.md"
---

# OpenAI says it reached an agreement with the Department of War to deploy advanced AI systems in classified

## Retrieval answer

OpenAI says it reached an agreement with the Department of War to deploy advanced AI systems in classified environments, with a cloud-only architecture, retained safety controls and human oversight.

## Summary

OpenAI says it reached an agreement with the Department of War to deploy advanced AI systems in classified environments, with a cloud-only architecture, retained safety controls and human oversight. A March 2 update explicitly bars intentional domestic surveillance of U.S. persons. For builders, the agreement is a concrete, consequential example of technical and contractual controls being combined in a sensitive deployment.

## What happened

OpenAI’s announcement, dated February 28, 2026, says it reached an agreement with the Department of War “yesterday” to deploy advanced AI systems in classified environments. OpenAI says it asked that the same terms be made available to all AI companies. The announcement describes three red lines: no use of its technology for mass domestic surveillance, to direct autonomous weapons systems, or to make high-stakes automated decisions that require human approval.

The company says deployment will be cloud-only. It will retain full discretion over the safety stack, including the ability to run and update classifiers, and says it will not provide “guardrails off” or non-safety-trained models. It also says it will not deploy models on edge devices, which it argues could enable use in autonomous lethal weapons. Cleared OpenAI engineers will be forward-deployed to help the government, with cleared safety and alignment researchers in the loop.

The agreement permits use for lawful purposes, subject to applicable law, operational requirements and safety and oversight protocols. Its quoted terms say the system will not independently direct autonomous weapons where law, regulation or Department policy requires human control, or assume other high-stakes decisions requiring human approval. They also refer to verification, validation and testing for AI use in autonomous and semi-autonomous systems. For intelligence activities, the quoted terms require handling private information to comply with specified U.S. laws and a defined foreign-intelligence purpose, and restrict unconstrained monitoring of U.S. persons’ private information.

In an update dated March 2, OpenAI says the Department and the company added explicit language that its tools will not be used for domestic surveillance of U.S. persons, including through commercially acquired personal or identifiable information. The Department also affirmed that its intelligence agencies, such as the NSA, will not use OpenAI’s services under this agreement; services to those agencies would require a new agreement. OpenAI says the Department plans to convene a working group involving frontier AI labs, cloud providers and government policy and operational communities, and that it will participate. The announcement gives no price, deployment timetable or model names.

## What the archive adds

New Runtime’s earlier coverage of [OpenAI’s Daybreak Red Team and Daybreak Defense Factory](https://newruntime.com/posts/openai-launches-daybreak-red-team-and-daybreak-defense-97701d/) on November 14, 2026, described an agent-based penetration-testing product and its early-access interest process. That was a separate cybersecurity offering; this agreement is a new event concerning classified government deployment and its stated safeguards. The shared thread is OpenAI’s attempt to define how advanced capabilities are accessed and constrained in security-sensitive work.

## Trend signal

This agreement supports the view that production-agent security boundaries are being defined partly by the runtime and deployment environment, not by behavioral instructions alone. OpenAI points to a cloud-only deployment, its own safety stack and classifiers, cleared personnel, and contract language as layered controls. Those are claims about its approach, not independent proof that the controls work as intended. The case is relevant to [agent security moving to runtime boundaries](https://newruntime.com/patterns/agent-security-moves-to-runtime-boundaries/).

## Connections to other events

OpenAI’s agreement follows a dispute with Anthropic over classified deployment terms, which OpenAI discusses in its own announcement. The company says it does not know why Anthropic could not reach a deal and argues its own contract provides stronger safeguards than Anthropic’s original contract. Separately, New Runtime’s November 14, 2026 coverage of Daybreak described an agent-based security product. These are distinct developments, but together they show OpenAI addressing security both through a product for testing web applications and through terms and controls for government deployment.

## Why it matters

For teams deploying AI in sensitive settings, the announcement illustrates that a policy statement is only one part of a control system. OpenAI describes restrictions expressed in contract language alongside deployment limits, a vendor-operated safety stack, ongoing classifier updates and cleared staff involvement. That combination is useful as a design question for any organization procuring systems whose outputs could affect people or critical operations: which restrictions are enforceable through the architecture, which depend on a human decision, and who can observe and intervene?

The announcement does not establish that this particular model or deployment will be used by other organizations, nor that the same controls are appropriate for every use. But it raises a practical procurement standard: request a description of where the system runs, what safeguards the supplier retains, how sensitive use is monitored, and what the contract requires when human control is legally mandated.

## Limits and uncertainty

OpenAI is describing its own agreement and its own safeguards; the announcement does not independently verify their operation or effectiveness. It does not identify the models, disclose a deployment schedule, provide performance or audit results, or explain how often classifiers will be updated. The public description also does not specify the technical process for detecting violations, the evidence a customer or outside reviewer could inspect, or how disputes over a suspected breach would be resolved. OpenAI says it could terminate the contract if the government violated its terms, but gives no process or example for doing so. The March 2 language narrows stated permissions, while implementation remains to be demonstrated.

## Skill to build now

For an AI procurement review, turn broad red lines into a control map. For each prohibited use, ask the supplier to identify the corresponding deployment restriction, safety control, human approval point, monitoring evidence and contractual remedy. Then mark which controls the buyer can verify and which depend on the supplier’s assurance. OpenAI’s announcement offers concrete categories for this exercise: cloud versus edge deployment, retained classifiers, cleared staff in the loop and explicit human-control terms.

## Reader takeaway

The agreement’s significance is not simply that a frontier model provider is working with the military. OpenAI says it is pairing contract restrictions with limits on where and how its systems run, retained safety controls and cleared personnel. Those are meaningful design choices, but their practical force depends on implementation and verification—details the announcement does not yet establish.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Anthropic](https://newruntime.com/topics/anthropic.md): Explore the anthropic topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
