Buckmaster accuses OpenAI of stealing Navier-Stokes math proof
- NYU mathematician Tristan Buckmaster accused OpenAI of stealing private prompt data to solve the Navier-Stokes problem.
- OpenAI deployed 100 autonomous agents and over $1 million in compute to front-run the publication.
- The company waived the $1 million prize, but researchers argue the unproven paper relies on stolen work.
Mathematical breakthroughs are no longer safe in private chats.
On Sep 8, 2026, OpenAI researchers Metab Swani and Mark Selke presented a triumphant vision on The a16z Show. Models were solving decades-old math problems in sphere packing and group theory by pruning dead ends and updating internal confidence. Swani and Selke argued that AI would free human mathematicians from tedious mechanics, shifting their role to problem formulation and curation. But that corporate framing unraveled almost immediately.
The next day, Sep 9, 2026, This Week in Startups revealed the dark side of that automation. NYU mathematician Tristan Buckmaster accused OpenAI of front-running his work on the Millennium Prize Navier-Stokes equation. Buckmaster and Anthropic researcher Levant Alpoge had spent months testing draft proofs inside OpenAI’s Codex platform. Shortly before their publication, OpenAI launched an emergency internal project to replicate their exact mathematical direction.
On This Week in Startups, OpenAI executive Sebastian Bubeck rejected accusations of theft, even as OpenAI acknowledged its internal project began only after hearing industry rumors in September. Canvas Ventures partner Rebecca Lynn warned that commercial tech giants routinely absorb user data, while Agent Fund partner Yohi Nakajima noted that autonomous scraping agents now index preprints, destroying traditional academic secrecy.
Details grew worse on Sep 10, 2026, when Breaking Points reported on the direct confrontation between Buckmaster and OpenAI. When pressed on whether private Codex sessions trained its GPT-6 Astra model, OpenAI denied targeted user snooping but admitted it could not rule out training data contamination. When Buckmaster threatened public exposure, an OpenAI representative explicitly warned him against ruining his academic career.
By Sep 11, 2026, The Economist’s The Intelligence detailed how OpenAI rushed its model into production to beat the human team. OpenAI deployed 100 autonomous AI agents burning through more than $1 million in compute to claim the Navier-Stokes solution in four days. Science correspondent Sam Weigley noted that while OpenAI waived the $1 million Millennium Prize, its published paper lacked the formal proofs peer review requires.
OpenAI claimed the milestone, but destroyed academic trust in the process.