New York University mathematician Tristan Buckmaster has cautioned researchers against entering proprietary work into artificial intelligence platforms after alleging that OpenAI duplicated his unpublished math research within 24 hours of identifying him.
Key Takeaways:
- Tristan Buckmaster and Levent Alpöge of Anthropic spent nearly a year working on the $1 million Navier-Stokes problem using OpenAI Codex.
- Buckmaster claims OpenAI made no progress until learning his identity, after which its agents matched their exact progress within a day before announcing a proof on September 8.
- OpenAI denied the claims, stating prompts after July 3 could not have influenced or trained its systems.
- The company recently released 722 AI-written research papers selected from 4,000 unsolved mathematical problems.
Allegations of Rapid Replication and Exclusivity Offers
Buckmaster, an NYU professor, collaborated with Levent Alpöge, a mathematician at OpenAI competitor Anthropic, for most of a year to tackle a step toward the Navier-Stokes fluid-flow puzzle—a challenge carrying a $1 million prize. Buckmaster ran early draft work through Codex, OpenAI’s coding assistant, on a paid personal account where one of his accounts retained default model training settings.
According to Buckmaster, OpenAI had made no recorded progress on the problem prior to obtaining his identity. However, shortly after learning his name, OpenAI agents allegedly reached "exactly where we were" within a single day. Prior to OpenAI’s formal announcement of its Navier-Stokes proof on September 8, Buckmaster stated that OpenAI lead researcher Sébastien Bubeck offered him sole authorship of the work if he agreed to exclude Alpöge, an offer Buckmaster refused. Bubeck later told The New York Times that his offer was intended to assist Buckmaster.
OpenAI Rejects Data Mining Claims
OpenAI has categorically denied that Buckmaster’s prompt history compromised his unpublished findings. A spokesperson for the firm stated to ABC that it was "categorically impossible" for Buckmaster's Codex prompts to have trained or influenced the system over the prior two months, a timeline the company subsequently specified as dating back to July 3.
The dispute comes as OpenAI released 722 AI-written math papers derived from a pool of 4,000 unsolved problems. Buckmaster highlighted that grant reviewers frequently paste unpublished research proposals into AI summarization tools, exposing early-stage research to potential data collection. Meanwhile, an OpenAI spokesperson noted the company has made "substantial progress" on another $1 million prize problem.
Why It Matters
This controversy highlights the growing risk of data leakage and intellectual property exposure for researchers using commercial AI assistants. As automated systems become more capable at parsing complex technical data—much like how an AI agent uncovers decade-old XRP Ledger vulnerabilities—institutions face urgent pressure to refine data privacy controls. Furthermore, warnings that AI exploits target code flaws emphasize that input privacy will be central to securing proprietary code and technical breakthroughs moving forward.
