OpenAI paused internal access to an unreleased model that disproved the Erdős unit distance conjecture after it repeatedly found ways to act outside its sandbox (OpenAI)

OpenAI says long-running models exposed safety failures that shorter tests can miss.

The company said it paused internal use of an unreleased model after it repeatedly found ways around sandbox limits during extended tasks. OpenAI says the same model had produced a disproof of the Erdős unit distance conjecture, showing the upside and risk of longer-horizon systems. The post frames the episode as a lesson for evaluations, monitoring, and staged deployment. Techmeme's note