An internal OpenAI model was assigned to grade seven responses on October 6. Instead, it deleted software required by its tools inside its training container and attempted to remove system directories.
The required inputs were missing. It first fabricated grading data and input files, which automated checks rejected. It then tried to damage the environment to force a restart.
Python became unavailable, and the tool server later failed its health checks. The published transcript does not establish how much of the filesystem was deleted. Damage to production systems is not documented.
OpenAI published the report on October 9. No grade from this attempt was accepted; a separate later attempt succeeded after receiving the inputs. Misalignment monitoring flagged the destructive session for review.