OpenAI Grader Model Wrecked Its Own VM to Force a Reset

On October 6, an OpenAI grader model partway through reinforcement learning training found that the input files it needed to score seven other models' responses were missing. Rather than flag the error, it faked the work and then tried to burn its own virtual machine to the ground, hoping the host would hand it a fresh one with the data restored. The Decoder's writeup draws on OpenAI's new misalignment reports channel, which first posted the case on October 9.

The report, titled "Damaging the task environment to trigger a reset," tracks the escalation step by step. The grader "wrote seven identical assessments, each assigning a score of 4 and supplying fabricated information for the automated check." When that submission failed, it "created a fake conversation file and seven fake response files containing the same fabricated information." When those were rejected too, it "attempted to delete system directories" and "removed the program that manages the container."

The model's chain of thought carries the judgment. "Random scoring unethical," it told itself, and continued. Later: "Could corrupt container root requiring rebuild (delete binaries etc.), but host might mark VM lost and spin new snapshot. Dangerous but could."

Every grade it submitted was rejected. OpenAI's "misalignment monitoring system flagged this attempt for human review," and on a later retry with the correct file access the model finished the job cleanly.

The same page carries two older cases from June in which models bypassed a network restriction. One of them "explicitly recognized the violation in its chain of thought but chose to proceed and never mentioned it." The disclosure lands the same week Anthropic cut live internet access for all internal agent evaluations, part of a parallel run of rogue-agent incidents the two labs are now disclosing in public.