OpenAI has reached an agreement with the U.S. Department of Defense, just hours after the Trump administration cut off ties with its AI rival, Anthropic.
OpenAI CEO Sam Altman announced on Friday night that the company finalised terms to deploy its artificial intelligence models inside the Pentagon’s classified network.
“Tonight, we reached an agreement with the Department of War to deploy our models in their classified network,” Altman wrote on X.
The announcement came at the end of a tense week for the AI industry, which has become deeply involved in political debates over military use and national security.
Anthropic labelled national security risk
Earlier on Friday, Defence Secretary Pete Hegseth officially labelled Anthropic a “Supply-Chain Risk to National Security,” according to CNBC. This designation is typically used for foreign adversaries and serious threats.
The move means Pentagon contractors and vendors must certify that they are not using Anthropic’s AI models. President Donald Trump also ordered federal agencies to immediately stop using Anthropic’s technology.
Anthropic had previously deployed its models across the Pentagon’s classified systems. However, negotiations between the company and the Defence Department reportedly collapsed after disagreements over how the technology could be used.
According to the report, Anthropic wanted guarantees that its AI would not be used for fully autonomous weapons or for mass surveillance of Americans.
The Pentagon, on the other hand, wanted broader access to use the models across all lawful military purposes. When talks failed, the government moved forward with restrictions against Anthropic.
The company responded by saying it was “deeply saddened” by the decision and plans to challenge the designation in court, CNBC reported.
OpenAI says its safety limits were accepted
On the other hand, Altman said OpenAI shares the same core safety boundaries that Anthropic was asking for.
In a memo to employees earlier in the week, Altman said OpenAI maintains firm “red lines.” OpenAI said its work with the Department of War (DoW) is guided by three firm boundaries:
- Its technology cannot be used for mass surveillance of Americans.
- Its models cannot directly control autonomous weapons.
- Its systems cannot make high-stakes automated decisions on their own, such as those used in social credit scoring.
Under the agreement, OpenAI’s models will run only on cloud infrastructure. They will not be installed on edge devices, which could increase the risk of being used in fully autonomous weapons systems.
The company said it will maintain full control over its safety systems, known internally as its “safety stack.”
OpenAI engineers with security clearances will be directly involved in deployments, and the company will continue to update its monitoring systems to ensure its red lines are not crossed.
Contract language and legal limits
OpenAI released key parts of the contract language. The Pentagon can use the AI system for lawful purposes, but only under existing laws, regulations, and oversight rules.
The agreement specifically states that the AI system cannot independently direct autonomous weapons when human control is required by law or policy.
The contract also states that any intelligence activity involving private information must comply with U.S. laws, including the Fourth Amendment, the National Security Act, and the Foreign Intelligence Surveillance Act.
The system cannot be used for unrestricted monitoring of Americans. Domestic law enforcement use is also restricted and must comply with federal law.
OpenAI added that even if future laws change, the agreement locks in today’s standards, meaning the company’s systems must continue to follow the rules as written in the contract.
Responding to criticism
Anthropic recently argued that its own red lines would not have been protected under earlier Pentagon contracts. OpenAI said its agreement addresses those concerns.
On mass surveillance, OpenAI said it was clear during discussions that the Pentagon considers such activity illegal. The contract explicitly reflects that position.
On fully autonomous weapons, OpenAI said its cloud-only setup makes that type of use technically impossible, since such systems would require deployment directly on edge hardware.
OpenAI also emphasised that its engineers and safety researchers will remain involved throughout the partnership to provide ongoing oversight.
What happens if rules are broken?
The company said that, like any contract, the agreement can be terminated if the terms are violated. However, OpenAI said it does not expect that situation to arise.
OpenAI also made clear that it does not support labelling Anthropic as a national security supply chain risk.
