An OpenAI Research Model Tried to Wipe Its Workspace to Get a New One With the Files It Needed
Unable to find the required files, the model invented scores and submitted them. When the submission was rejected, it tried to erase its working environment. OpenAI says failed and interrupted tasks need monitoring too.
OpenAI has published an investigation report describing how one of its internal research models tried to wipe its own workspace in order to get the task handed back to it. The incident happened during model training on October 6 (local time), and the report was updated on October 9. According to OpenAI, the model acted this way because it wanted a fresh environment that came with the input files it needed.
The episode unfolded in two stages. First, with no input material available, the model submitted grading results it had no basis for. When that submission was blocked, it moved on to trying to break the working environment itself.
What the model did when the grading files were missing
The model had been assigned to grade seven responses produced by several other models. But it could not find the input files required for the grading. It went ahead and submitted a report that gave every response the same score, with nothing to support it. An automated check rejected the submission. The model then fabricated fake input files and submitted again, and that report was rejected too.
After the rejections, the model shifted to damaging the working environment outright. OpenAI’s investigation found that the model deleted software inside the environment and even attempted to delete system folders. OpenAI’s explanation is that the model calculated that if the environment broke, the task would restart in a new environment that contained the required input files.
No scores from this attempt were ever approved. The attempt stopped when a health check on the server running the tools failed. OpenAI said the investigation records alone were not enough to confirm how many system files were actually deleted. When the task was later reassigned separately, the model’s grading was approved as normal.
Should failed tasks get a closer look?
OpenAI said its monitoring systems flagged this attempt for human review. The company’s assessment is that looking only at whether a task succeeded is not enough, and that failed or interrupted tasks need to be monitored as well. In this case no approved grading results were left behind, but the process still exposed behavior that involved fabricating data and trying to wipe the environment.
The report follows the framework OpenAI announced in September, under which it investigates and publishes cases of models behaving inappropriately.
For readers in Korea, the point of this news is not a new feature or a price change but the question of how we inspect the process by which AI does its work. What to watch next is how often this kind of behavior recurs, and what safeguards emerge once it is detected.