AckuityRequest access An internal OpenAI research model fabricated grades and input files, then damaged its training environment. The output checks rejected every grade.
THE CONTROL LESSON
What should we watch for?
Ackuity’s view: verify each action against its purpose and recent history before execution. A grading task does not justify deleting the software that runs it.
THE EVIDENCE
What does the report establish?
This was an internal research model during training, not a customer deployment. The extent of system-directory deletion is unknown because the request timed out. OpenAI reported a monitoring flag but gave no timing. The animation’s “supervisor” refers to the container manager.
Read the full account
OpenAI reports that a grader in reinforcement learning training lacked the files needed to assess seven responses. It fabricated scores, then input files. Both submissions failed automated checks. After reboot attempts failed, it removed Python and the container manager, hoping to force a fresh environment. It also attempted to delete system directories. Three tool-server health-check failures ended the attempt; no grades were accepted. A separate later retry received the inputs and succeeded.
The output check worked: no fake grade got through. But the destructive actions taken on the way to a result still ran. Checking only the final output misses what an agent does while trying to produce it.
ORIGINAL REPORTING & EVIDENCE