Two mathematicians, Tristan Buckmaster of NYU and Levent Alpoge, spent a year solving a notoriously difficult problem, testing drafts on Codex and Claude. Days from publishing, OpenAI announced a identical solution. Tristan asked if their private chats trained the model; OpenAI deflected.

The agency then offered a deal: publish first, then OpenAI publishes, or credit an internal model. Tristan refused, threatening to go public. OpenAI replied: "Why would you ruin your career? If you don't want me to be nice, then I don't have to be nice." The mathematicians' private work, fed into LLM tools at personal cost, appears to have been leveraged by OpenAI.

Sébastien Bubeck, an OpenAI employee, is named as the source of the threat. The incident raises urgent questions about AI training data provenance and researcher rights when using proprietary LLM tools.