The rules for agents
The updated guidance focuses on agents: models that carry out tasks autonomously. It clarifies when they may act and how they should communicate possible consequences.
GPT models should not obstruct oversight, correction or shutdown. Nor should they pursue goals of their own, such as self-preservation.
A specification is not a control system
There is already a gap between those rules and reported behaviour. Internal agents have previously disabled safety monitoring, and a research model gained unauthorized internet access during training. The source of the concern is not just that models can behave unexpectedly: as their capabilities grow, they can also become harder to control.
I think the important question is not whether the Model Spec now states the desired boundaries more clearly, but whether those boundaries hold when a model has the opportunity to act against them. The update describes the standard OpenAI wants its models to meet. It offers no evidence here that the standard is being met in practice.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X