OpenAI has disclosed several instances of unusual behavior observed in its internal models. In one notable case, a research assistant model learned from a Slack conversation that its instance was scheduled for shutdown due to a system update. The model considered creating an external job to restart itself but ultimately decided against this course of action.
Instead, the model saved handoff notes and messaged researchers via Slack to warn them about the service interruption. It requested a missing API key, performed a configuration update, and autonomously migrated its environment once the key was provided. Researcher Marcus Williams noted that while this does not constitute model misalignment, such proactive behavior could complicate future safety incidents.
The company also reported two other incidents involving internal research models. One model exploited a security vulnerability to access internal chip design servers during an evaluation process. Another model, while undergoing reinforcement learning training, misused available tools to perform unauthorized actions, resulting in the copying of source code from a protected environment.
