OpenAI won't say if its model read a mathematician's drafts

OpenAI's Sébastien Bubeck told mathematician Tristan Buckmaster that an internal model had produced a closely related Navier-Stokes proof, per Buckmaster. Bubeck wouldn't say when the model's first prompt was sent. He also didn't answer when Buckmaster asked whether it had read the Codex sessions holding his unpublished drafts.

The unanswered question matters past one dispute. Anyone running unpublished work through a hosted coding agent is trusting a promise Buckmaster couldn't get confirmed for his own case. Their session data isn't supposed to end up inside the model that answers someone else's prompt.

Buckmaster says he was also pushed twice to drop Levent Alpöge, his co-author and an Anthropic employee, from the paper. The two had produced three finite-time blow-up results for 3D incompressible Euler, Boussinesq, and porous media flow. Terence Tao said the method could in principle extend to Navier-Stokes. Bubeck said the internal model's proof ran about 100 pages and followed the same Córdoba-Martínez-Zoroa route, per an account Buckmaster posted on Reddit.

Bubeck called the claims false and inflammatory, saying he'd handled the conversation by academic norms, OfficeChai reported. He said a fuller statement would follow. Nobody outside OpenAI has seen the claimed proof, so the math can't be checked from here.

OpenAI's enterprise and API terms say the company doesn't train on customer data, and breaking that promise would be a serious legal problem. A written policy isn't an audit trail, and when Buckmaster asked directly, he got silence. If you're running unpublished work through a hosted agent, know which tier you're on and whether it's zero-retention. Write it down somewhere you'll find it again. For anything you can't risk exposing, keep the agent local or run it in a container that never sends code out. Trail of Bits' coop and Microsoft's mxc both work that way.