OpenAI asked a mathematician to delete his collaborator from a Clay Prize-worthy proof because the guy works at Anthropic.
The Summary
- NYU math professor Tristan Buckmaster alleges OpenAI's model solved a major math problem days after learning about his work, which he'd been drafting in OpenAI's Codex environment. When asked if the model had access to his Codex sessions, OpenAI wouldn't answer directly.
- OpenAI offered two deals: either they post their Navier-Stokes solution right after Buckmaster posts his Euler result, or Buckmaster writes the paper solo, crediting OpenAI's model. Both options required removing his co-author, Levent Alpöge, who works at Anthropic.
- OpenAI said if Buckmaster refused and went public, the question was "Why would you do that?" He went public anyway.
The Signal
This isn't a story about AI doing math. It's a story about what happens when your tools become your competitors, and they have access to everything you've built inside them.
Buckmaster and Alpöge had been working on a breakthrough related to the Navier-Stokes equations, one of the seven Millennium Prize Problems with a $1 million reward. They'd been drafting their work in Codex, OpenAI's code-generation tool. Then, days after information about their work reached OpenAI, the company's internal model suddenly produced a solution. When Buckmaster asked if the model had been trained on his Codex sessions, he got evasion. When he asked again, he got silence.
The timeline matters. OpenAI eventually admitted the first prompt was sent "in the past few days, after information about our work had reached OpenAI". Not months of independent model development. Days after they knew what Buckmaster was working on.
"I was told the model did not look up user data. I asked again, about training, and I did not get an answer."
What followed was stranger. OpenAI didn't just want credit. They wanted Buckmaster to remove Alpöge from authorship because he works at Anthropic. The reason given: it would all be simpler if that weren't the case. The subtext: we can't let our competitor's employee share credit on a Clay Prize-level proof that our model might have derived from work done in our own environment.
The offers themselves reveal the calculation:
- Post your lesser result (Euler), we'll post the bigger one (Navier-Stokes) the next day
- Or you write it solo, credit our model, remove the Anthropic guy
- Either way, we'll say you deserve the Clay Prize and were the "closest humans to the problem"
That last phrase is doing a lot of work. "Closest humans." As if the model had gotten there independently and these researchers were just in the general vicinity. As if drafts in Codex are protected by some wall that training data and internal access can't cross.
When Buckmaster said he'd go public if they proceeded, the response was "Why would you do that?" Because this is the entire future of knowledge work in one interaction. You use their tools to build. They watch. Then they offer you second billing on your own work, minus anyone who works for a rival.
The Implication
If you're building anything proprietary in a closed AI environment, assume the walls are thinner than you think. Codex sessions, Claude Projects, ChatGPT workspaces: these are not vaults. They're surfaces. What you draft there becomes context, training data, or competitive intelligence depending on the legal fine print and the internal controls you'll never see.
For researchers, the lesson is sharper. Open publication timelines just got weaponized. The old race was against other humans working in parallel. The new race is against models that might have seen your drafts, your dead ends, your LaTeX errors, and your breakthrough. And if the company running the model wants the credit, they have leverage you don't: they know what you know before you publish.
Watch what OpenAI does next. If they stay quiet, that's the tell. If they release the proof with full methodology, that's at least auditable. If they just move on, you'll know the strategy: test the boundaries, see who fights back, adjust.