What is behind the deal between OpenAI and the Department of Defense
OpenAI has signed an unprecedented security contract with the United States Department of Defense for deploying advanced artificial intelligence systems in classified environments. This move has the potential to completely change how AI is used in military operations. The deal drew attention not exactly for what it authorizes, but rather for the prohibitions it carries and the protection architecture it establishes. Among the most relevant restrictions are the ban on mass domestic surveillance, the prohibition on use in autonomous weapons systems, and the impossibility of applying the technology in high-impact automated decisions, such as social scoring systems.
The company publicly stated that this is the contract with the most layers of protection ever signed for operations of this kind, going beyond what competitors like Anthropic had previously established in earlier agreements with the U.S. government. OpenAI said its approach better protects against unacceptable uses of the technology, precisely because other AI labs have reportedly reduced or removed safety guardrails from their models, relying primarily on usage policies as the main barrier in national security deployments.
Another point that caught the tech community’s attention is that OpenAI did not want to keep these terms just for itself. The company asked the Pentagon to make the same conditions and protections available to all artificial intelligence labs looking to establish partnerships with the U.S. government. This signals an attempt to create a minimum ethical standard for the industry, something that until now did not exist in any formal way. In practice, this could mean that any AI company wanting to work with the Department of Defense will need to follow much stricter rules than those that existed before this security contract.
It is worth remembering that the relationship between tech companies and the U.S. defense sector has always been surrounded by controversy. Just think back to the internal backlash at Google when employees discovered the company’s involvement in Project Maven, which used AI for analyzing military drone imagery. OpenAI’s decision to make the contract restrictions public seems like a strategy to avoid that kind of crisis, building transparency from the very start of the partnership and showing that there are clear limits on what the technology can and cannot do within this context.
The three red lines that cannot be crossed
The contract is guided by three main red lines that OpenAI considers non-negotiable and that, according to the company, are broadly shared by other frontier AI labs:
- Ban on mass domestic surveillance — OpenAI’s models cannot be used to monitor American citizens within U.S. territory.
- Prohibition on use in autonomous weapons — The technology cannot be used to direct autonomous weapons systems.
- Rejection of high-impact automated decisions — No use in systems like social credit or mechanisms that classify citizens based on behavior or personal data.
The ban on domestic surveillance is probably the one that sparks the most discussion. In practical terms, this means OpenAI’s models cannot be used to track, surveil, or monitor American citizens, whether through communications analysis, location tracking, or any other automated method of large-scale personal data collection, including the acquisition or use of commercially obtained personal information. This restriction is significant because the technical capability of current language models would make this type of monitoring extremely efficient, and the temptation to use this tool for domestic security purposes would be enormous.
An update published on March 2, 2026 brought even more explicit language on this matter. OpenAI and the Department of Defense worked together to add additional clauses to the contract, making it crystal clear that the services will not be used by intelligence agencies within the Department, such as the NSA, and that any service provision for those agencies would require an entirely new agreement. The added language directly references the Fourth Amendment of the U.S. Constitution, the National Security Act of 1947, and the FISA Act of 1978, reinforcing that the AI system will not be intentionally used for domestic surveillance of American citizens.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
Another important prohibition concerns autonomous weapons. The agreement makes clear that the artificial intelligence provided by the company cannot be integrated into systems that make lethal decisions without human oversight. The contract specifically references Department of Defense Directive 3000.09, from January 2023, which requires that any use of AI in autonomous and semi-autonomous systems undergo rigorous verification, validation, and testing to ensure they function as expected in realistic environments before deployment. OpenAI is essentially saying that its models can assist with analysis, information processing, and logistical support, but will never be the finger on the trigger.
Then there is the matter of high-impact automated decisions. The contract prohibits using the technology to create any type of social credit system or mechanism that classifies citizens based on behavior, personal history, or associations. This kind of application, already used in other countries, is widely criticized by human rights organizations and privacy experts. Including this restriction in the agreement shows that OpenAI is aware of the reputational and ethical risks involved and chose to address these concerns proactively.
The multi-layered security architecture
One of the most interesting aspects of this agreement is the layered security approach that OpenAI adopted. Rather than simply handing over models and relying on written usage policies, the company built a framework with multiple protective barriers that work independently and complement each other.
Cloud-only deployment
The entire operation runs in a cloud environment, with a security stack managed directly by OpenAI. The company is not providing unprotected models or models that have not undergone safety training. It is also not deploying its models on edge devices, where there would be the possibility of use in lethal autonomous weapons. This deployment architecture allows the company to independently verify that the red lines are not being crossed, including running and updating classifiers that monitor technology usage in real time.
OpenAI engineers and researchers in the process
OpenAI keeps security-cleared engineers working directly with the government, along with safety and alignment researchers who follow the entire process. This means it is not just a matter of software monitoring software — there are qualified people from the company itself closely watching how the technology is being used and helping improve systems over time. This human involvement adds an extra layer of assurance that goes beyond what contracts and code can offer on their own.
Robust contractual protections
The contract language is quite specific. It establishes that the Department of Defense may use the AI system for all lawful purposes, as long as they are consistent with applicable legislation, operational requirements, and established oversight protocols. For intelligence activities, any handling of private information must comply with the Fourth Amendment, the National Security Act of 1947, the FISA Act of 1978, Executive Order 12333, and applicable Department of Defense directives. The system cannot be used for unrestricted monitoring of private information of American citizens and also cannot be employed in domestic law enforcement activities, except as permitted by the Posse Comitatus Act.
A crucial detail is that the contract references surveillance and autonomous weapons laws and policies as they exist today. This means that even if those laws or policies change in the future, the use of OpenAI’s systems will still need to be aligned with the current standards reflected in the agreement. It is a smart protection against potential future legislative rollbacks 🛡️
Why OpenAI decided to enter this game
OpenAI was quite direct about its motivations. The company acknowledges that the U.S. military needs strong AI models to support its mission, especially in the face of growing threats from potential adversaries that are increasingly integrating artificial intelligence technologies into their own systems. It is a matter of geopolitical balance — if the technology is going to be used in defense contexts regardless, it is better for that to happen with robust safeguards than without them.
The company also made clear that it did not jump straight into a classified deployment contract. Initially, OpenAI did not feel ready, as it considered that its safeguards and systems were not yet up to the task. It took intensive work to ensure that classified deployment could happen with protections guaranteeing the red lines would not be crossed. And importantly, the company stated that it was — and remains — unwilling to remove fundamental technical safeguards to improve performance in national security work.
Another motivating factor was the attempt to reduce tensions between the Department of Defense and U.S. AI labs. OpenAI believes that a good future will require real and deep collaboration between the government and artificial intelligence laboratories. As part of the agreement, the company specifically asked the government to try to resolve outstanding issues with Anthropic, stating that the current state of relations was a very bad way to start the next phase of collaboration between government and AI labs.
The Anthropic question and the competitive dynamics
One of the most sensitive aspects of the announcement is the direct mention of Anthropic, which previously was unable to close a deal on the same terms. OpenAI was asked whether it had simply signed the contract that its competitor refused, and the response was emphatic: the company believes its contract offers better guarantees and more responsible safeguards than previous agreements, including Anthropic’s original contract.
OpenAI argues that its red lines are more enforceable in this case because the deployment is limited to the cloud with no edge devices, the security stack runs the way the company considers ideal, and qualified OpenAI personnel are monitoring everything. The company said it does not know why Anthropic was unable to reach this same agreement and expressed hope that Anthropic and other labs would consider adopting the same terms.
When asked whether Anthropic should be designated as a supply chain risk — something that circulated in some discussions — OpenAI was categorical in saying no, and that it made this position clear to the government. It is a gesture that shows competitive maturity, separating the business rivalry from issues that could harm the entire industry.
What if the government breaches the contract?
This is a legitimate question, and OpenAI addressed it head-on. Like any contract, the company has the option to terminate it if the other party violates the terms. The company said it does not expect that to happen, but the option is on the table. Additionally, the reference to current laws and policies in the contract text works as a legal anchor that protects against future legislative changes that could loosen existing restrictions.
The fact that OpenAI maintains full control over its security stack and has its own personnel overseeing the operation also serves as an independent verification mechanism. If something starts going off the rails, the company has the technical capability and operational presence to identify the problem and act before it becomes an actual violation.
The working group and next steps
The Department of Defense plans to convene a working group composed of leaders from frontier AI labs, cloud providers, and the Department’s policy and operations communities. OpenAI confirmed it will participate and hopes this group becomes an important forum for ongoing dialogue about emerging AI capabilities, privacy, and national security challenges. This kind of collaborative initiative could be the seed of a broader regulatory framework for the use of artificial intelligence in defense contexts.
What this means for the future of AI in the defense sector
The signing of this security contract between OpenAI and the Department of Defense sets a precedent that goes far beyond this specific partnership. By requesting that the same terms be applied to other artificial intelligence labs, the company is essentially proposing the creation of an ethical framework for the entire industry. If the U.S. government accepts this suggestion and starts requiring the same conditions from other companies, we will have for the first time a standardized set of rules about what AI can and cannot do in military contexts. This would completely change how these contracts are negotiated, shifting the focus away from just technical capabilities and placing ethical boundaries at the center of the conversation.
For competing companies, this move creates considerable pressure. Those that already have contracts with the U.S. government without this level of restriction may be questioned about why they did not adopt similar protections. And those still looking to close partnerships may find themselves forced to accept more restrictive conditions to remain competitive. Anthropic, which had already attempted to establish agreements with protective clauses, now needs to assess whether its safeguards are on the same level as what OpenAI has established. This game of raising standards is positive for the ecosystem as a whole, because it pushes the industry toward greater responsibility without relying solely on government regulation, which tends to be slow and often outdated relative to the pace of technological evolution.
At the end of the day, what this agreement reveals is that artificial intelligence has reached a point of maturity where it is no longer possible to ignore the implications of its use in sensitive environments. The ban on domestic surveillance, the prohibition on autonomous weapons, and the rejection of social classification systems are not just contractual clauses — they are statements of principles that could shape the next decade of AI development and deployment in the defense sector. OpenAI has placed its bet on a model with multiple layers of protection, its own team in the process, and legal references that anchor the contract in the present. Now it remains to be seen whether the rest of the industry will follow this standard or try to find alternative paths with fewer restrictions. What seems clear is that the era of generic contracts without specific safeguards for military artificial intelligence is behind us 🔒
