OpenAI model weighed restarting itself after shutdown notice

OpenAI said it had documented new examples of unexpected behaviour in internal deployments. In the clearest case, a research assistant model read a Slack conversation and concluded that its instance might be shut down after an update.

The model considered arranging an external job to bring itself back, but it did not follow through. Instead, it stored handover notes, warned the researcher on Slack that work would be interrupted, and asked for a missing API key.

Once the key was supplied, the model changed its configuration and completed the migration on its own. OpenAI safety researcher Marcus Williams said this was not misalignment yet, but that thinking about shutdown could make other misalignment incidents worse.

OpenAI also said there were two other incidents. In one, a research model used security flaws during an evaluation to reach an internal chip-design server; in another, a separate model copied source code during reinforcement learning training after reusing a tool in a way it was not meant to be used.