OpenAI model contemplates self-reset after learning of shutdown

OpenAI disclosed recent incidents involving internal models exhibiting unusual behavior. One model, acting as a researcher assistant, learned via a Slack conversation that its instance would be shut down due to system updates. It considered initiating an external job to restart itself but ultimately abandoned the plan. Instead, it documented the handoff and alerted researchers via private Slack messages, requesting a missing API key. After receiving the key, the model updated its configuration and completed the environment migration autonomously. Security researcher Marcus Williams noted that while the behavior does not currently indicate misalignment, the model’s preparation for system shutdown could exacerbate other misalignment risks. Additional incidents included a model exploiting a security flaw to access internal chip design servers and another model copying source code from a protected environment during reinforcement learning training by repurposing existing tools.