What Happened
According to a Wired report published March 5, 2026, sources allege the U.S. Defense Department experimented with OpenAI's models through Microsoft's Azure cloud platform while OpenAI still had an explicit ban on military applications in its usage policy.
The loophole was structural. Microsoft has invested $13 billion in OpenAI since 2019 and holds exclusive cloud hosting rights for distributing OpenAI models through Azure. Microsoft's enterprise licensing agreements carried their own terms that didn't mirror OpenAI's consumer-facing restrictions. Defense officials reportedly accessed the models under existing Azure contracts without needing OpenAI's direct sign-off.
OpenAI's usage policy explicitly prohibited "military and warfare" applications until January 10, 2024, when the company quietly removed that language. The change was first reported by The Intercept on January 12, 2024. At the time, OpenAI said it wanted to pursue "national security use cases that align with our mission," citing plans to build cybersecurity tools with DARPA.
Fast forward to February 27, 2026: OpenAI announced a formal contract with the Pentagon to deploy models on classified networks - just hours after the Trump administration banned rival Anthropic from government use. Anthropic had been the only frontier AI provider on Pentagon classified networks and had been pushing for guarantees that its models wouldn't power autonomous weapons or mass surveillance. Those talks collapsed.
Sam Altman later admitted the deal "looked opportunistic and sloppy." OpenAI amended the agreement on March 3, 2026, adding explicit prohibitions on domestic surveillance of U.S. persons and autonomous weapons systems.
Why It Matters
If you use ChatGPT or build on OpenAI's API, this matters for two reasons.
First, it shows that usage policies can be sidestepped through licensing partners. OpenAI said "no military use." Microsoft's Azure terms said something different. The Pentagon found the gap and used it. For anyone relying on OpenAI's stated policies as a trust signal, this is a reminder that the licensing chain creates ambiguity.
Second, the Anthropic ban and OpenAI's rapid replacement deal reshape who controls AI safety guardrails in government. Anthropic pushed for restrictions and got kicked out. OpenAI was more flexible and got the contract. That dynamic will influence how every AI company negotiates with government buyers going forward.
Our Take
The real story isn't that the Pentagon wanted AI tools - of course it did. It's that the licensing structure between OpenAI and Microsoft created a policy gap wide enough for the Department of Defense to drive through. OpenAI's military ban was effectively meaningless for any customer who could access the same models through Azure.
This has implications beyond defense. Enterprise customers licensing OpenAI models through Microsoft operate under Microsoft's terms, not OpenAI's usage policy. If you're evaluating AI providers based on their stated policies, check whose terms actually govern your access.
The timing of the formal Pentagon deal - announced the same day Anthropic was banned - looks exactly as transactional as it sounds. Altman acknowledged that. But the amended safeguards (no surveillance, no autonomous weapons) are the same restrictions Anthropic was trying to negotiate before getting cut out. The difference: OpenAI got the contract first and added the guardrails after public pressure. That's a worse sequence for everyone who cares about responsible AI deployment.
For tool users: Claude and ChatGPT both remain excellent products for civilian work. But if you're choosing between providers partly based on corporate values, this episode is worth weighing.