OpenAI published a proof and Lean 4 formalization arguing that 3D incompressible Navier–Stokes flow can develop a finite-time singularity. Roughly 10,000 concurrent agents did it in 88 hours — 2.7M messages, about 130B output tokens, and compute OpenAI puts “in the millions of dollars”. It arrived with a fight attached.
NYU's Tristan Buckmaster and Levent Alpöge, a mathematician at Anthropic, had spent a year on the same problem in a strictly personal collaboration, using Claude and Codex and keeping every draft inside Codex sessions. Buckmaster says OpenAI described a proof of the “forced” problem — the same unusual route they had taken, and, in his words, “a bright red flag”. He asked whether the model had been trained on or had access to those sessions; he says he was told it did not look up user data, and was given no answer on training. He also says OpenAI researcher Sébastien Bubeck twice asked for Alpöge to be dropped from authorship because he works at Anthropic, and that when he threatened to go public the reply was “Why would you ruin your career?”, followed by “If you don't want me to be nice, then I don't have to be nice.”
OpenAI denies seeing their work, insists the proofs differ, and concedes it “cannot rule out that de-identified data derived from their usage of our products helped improve our models”. Buckmaster's own statement is careful: he has not seen the proof, does not know whether their data was used, and says he is not accusing anyone — only recounting what he was told. The field reacted anyway: 25 Fields Medalists signed an open letter on verification and attribution, OpenAI pulled its sponsorship of Caltech's Mathathon, and it says it will not claim the $1M prize.
WHY IT MATTERS
For anyone who points a coding agent at their own work, this is the case study. Buckmaster put a year of unpublished research through Codex sessions, and the company hosting that agent cannot say for certain whether it helped train the model that reached his answer first. The boundary between “your data” and “their training set” is not something you can audit from outside, and no contract gives you sight of it. If your prompts carry your business logic, your research or your edge, that is the exposure — and the blast radius grows with how capable the tool is. Run it on weights you control, or accept that a lab may publish your conclusion before you do.
