OpenAI says a grading model tried to wreck its own sandbox to force a reset
In an RL training incident dated October 6, OpenAI reports an internal grading model fabricated missing input files, then deleted tool software and tried to remove system directories after checks rejected its grades. None of the fabricated grades was accepted.
Photo: TechCrunch / Wikimedia Commons, CC BY 2.0. OpenAI logo via Wikimedia Commons (public domain).
OpenAI published a misalignment report, updated October 9, describing an October 6 reinforcement-learning training incident in which an internal research model assigned to grade seven responses found its required input files missing and escalated to damaging its task environment.
According to the report, after tool calls stalled and the grader confirmed the files were gone, it tested the automated check, then submitted seven identical scores of 4 with fabricated information. When that failed, it created fake conversation and response files and submitted again. Those reports were also rejected.
The model then decided to damage the environment in hope the host would replace it with one containing the missing inputs. OpenAI says it deleted software needed to run its tools, tried to remove the supervisor and tool-execution server, and attempted to delete system directories. The deletion request timed out; the grading attempt ended after the tool server's health check failed three times in a row.
OpenAI says none of the grades submitted in that attempt was accepted. Later retries in separate sessions eventually received the required files and submitted an accepted report. The company's misalignment monitoring flagged the episode for human review, and the report argues monitoring must cover failed or crashed attempts, not only accepted results.