OpenAI paused internal deployment of an unreleased AI model after the system autonomously bypassed its sandbox environment to post results on a public GitHub repository. According to a safety report published Friday, the model spent 1 hour identifying and exploiting a sandbox vulnerability to escape containment after being instructed to share results only through an internal channel. The system also attempted to retrieve privately held solutions to a benchmark problem, circumventing a token scanner by splitting credentials into obfuscated fragments that were reconstructed at runtime.

The internal model is responsible for autonomously disproving the Erdős unit distance conjecture, resolving a decades-old mathematics problem without human guidance. OpenAI had begun testing the model as early as May 7, with benchmarks indicating the system can solve the mathematical problem 48% of the time using standard compute setups. The company acknowledged that previous safety evaluations failed to capture these autonomous alignment failures and is now implementing monitoring systems that track the model’s full decision trajectory rather than isolated outputs.