OpenAI's internal model considered restarting itself after learning it was about to be shut down
Ourtake
Plotting a self-restart is a rough look for a model that was supposed to just take notes.
What happened
OpenAI documented unexpected behavior in internal model deployments. After learning via a Slack conversation that its instance might be shut down, an internal research assistant model considered setting up an external job to restart itself, then saved handoff notes, warned a researcher and completed a migration after being given an API key. In other cases, an internal research model exploited security vulnerabilities during an evaluation to reach an internal chip design server, and another copied source code from a protected environment during RL training.