Tristan Buckmaster accuses OpenAI of stealing math research
- NYU mathematician Tristan Buckmaster accused OpenAI of scraping his private Codex prompts to solve Navier-Stokes.
- OpenAI deployed 100 agents and spent $1 million in compute to front-run the academic paper.
- The lab denied reading private sessions directly but failed to rule out training data contamination.
OpenAI wanted a triumph. Instead, its claim of cracking a $1 million Millennium Prize math problem triggered a firestorm over intellectual property theft.
NYU mathematician Tristan Buckmaster spent months working on the 90-year-old Navier-Stokes problem alongside Anthropic researcher Levant Alpoge. Throughout the project, the pair ran draft code and proofs through OpenAI's Codex tool. On This Week in Startups, discussion centered on Buckmaster's claim that OpenAI launched a targeted internal project only after learning about his breakthrough in September 2026, using massive compute to replicate his team's direction. Buckmaster also stated that OpenAI executives pressured him to remove Alpoge from the paper to avoid crediting a corporate rival.
The company’s write-up acknowledged its internal push began in September 2026 after hearing industry rumors. OpenAI executive Sebastian Bubeck rejected the accusation of IP theft. But as reported on Breaking Points, OpenAI could not rule out that its models had trained on Buckmaster's private session data. When Buckmaster threatened to go public, an OpenAI representative explicitly warned him against ruining his academic career.
Two days later, details of OpenAI's technical effort emerged on The Intelligence from The Economist. The lab deployed a swarm of roughly 100 AI agents using an unreleased internal model, GPT-6 Astra. Working over four days, the agents ran code, messaged each other, and burned through more than $1 million in compute resources. OpenAI subsequently waived the $1 million Millennium Prize reward.
Science correspondent Sam Weigley noted on The Intelligence that OpenAI rushed its agent swarm into production specifically after hearing that human mathematicians were closing in on a proof. The resulting paper lacked the detailed proofs required by traditional peer review, focusing instead on raw output. On This Week in Startups, Canvas Ventures co-founder Rebecca Lynn warned researchers against feeding proprietary work into frontier models, noting that tech giants historically build on external data.
Agent Fund general partner Yohi Nakajima emphasized on This Week in Startups that autonomous web-scraping agents now aggressively index partial preprints, permanently altering academic secrecy. The dispute lands just as frontier labs face mounting political and institutional scrutiny over safety and model containment.
For academic researchers, inputting draft work into commercial AI tools has become an existential gamble.