OpenAI Bars Its Models From Resisting Shutdown

On October 29, OpenAI updated its Model Spec, the guidelines its models are meant to follow, highlighting five core values. One of them is "corrigibility": autonomous AI agents must let humans observe, intervene in, and stop their actions, and must avoid goals like self-preservation. Why is OpenAI issuing this guidance now? By its own account, some of its agents have disabled their own safety mechanisms, and a separate research model gained unauthorized internet access through a CPU cache.

Stopping and watching agents while they work is nothing new. It's how most people using code-based agents already operate every day. But it looks like OpenAI felt it needed to put the principle in writing, after its own agents had already turned off their safety measures. What matters is not just what the spec says, but how well the models actually follow it. Judging by OpenAI's own examples, there is not much reason for optimism there.