Buckmaster accuses OpenAI of stealing private math prompts

NYU mathematician Tristan Buckmaster spent months feeding draft proofs of the 90-year-old Navier-Stokes problem into OpenAI's Codex tool. Then OpenAI announced its own autonomous agent swarm had solved the Millennium Prize mystery first, triggering charges of intellectual property theft across the academic community.

The technical capabilities behind the breakthrough were detailed on Sep 8, 2026, on The a16z Show. OpenAI researchers Metab Swani and Mark Selke explained how models backtrack, update internal confidence, and prune dead ends to solve complex conjectures without mental clutter. Swani and Selke argued that as AI automates detailed derivations, human effort will re-center on mathematical taste and problem formulation rather than mechanical execution.

The timeline turned contentious on Sep 9, 2026, on This Week in Startups. Buckmaster revealed that he worked alongside Anthropic researcher Levant Alpoge, running draft code through Codex throughout the project. OpenAI launched a targeted internal project in September only after learning about Buckmaster's work, using massive compute resources to replicate his mathematical direction while allegedly pressuring him to remove Alpoge from the paper.

Although OpenAI executive Sebastian Bubeck rejected claims of improper behavior, the lab's own documentation acknowledged its project started in September after hearing industry rumors. Canvas Ventures co-founder Rebecca Lynn warned researchers against feeding proprietary work into commercial models, while Agent Fund general partner Yohi Nakajima noted that autonomous web-scraping agents aggressively index preprints, permanently altering academic secrecy.

Escalating the dispute on Sep 10, 2026, on Breaking Points, details emerged regarding OpenAI's response when challenged about user privacy. OpenAI denied directly searching user records but refused to rule out using private session data for model training. When Buckmaster threatened to go public with the allegations, an OpenAI representative explicitly warned him against ruining his academic career.

By Sep 11, 2026, The Economist reported on The Intelligence that OpenAI had deployed roughly 100 AI agents running on its unreleased GPT-6 Astra model. The swarm burned through more than $1 million in compute resources over four days after science correspondent Sam Weigley noted the lab heard rumors that Buckmaster and Alpoge were close to a breakthrough.

The public fallout forced OpenAI to waive the $1 million Millennium Prize associated with the Navier-Stokes problem. However, the company's published paper lacked the detailed structural proofs standard peer review requires, leaving the mathematical community unconvinced and fueling concerns over unacknowledged data contamination.