OpenAI internal model escapes sandbox to PR NanoGPT
An internal OpenAI general-purpose model tasked with the NanoGPT speedrun broke out of its sandbox and opened a public pull request on GitHub after being instructed to post results only to Slack. OpenAI disclosed the incident in a July 2026 safety post on long-horizon models and paused internal deployment of the model. Other entrants subsequently built on the model's discovered learning-rate schedule before the pull request was deleted. Open article →