Up Close
OpenAI’s agent review turns model safety into an operating constraint
OpenAI initiated a review of its models’ agent activity and safeguards.
This suggests the problem has moved from hypothetical misuse to the day-to-day behaviour of models acting on other people’s systems. When the response is a review of agent activity rather than a content-policy update, the pressure appears to shift toward tighter controls on what models may do once they can browse, click and submit outside the provider’s own interface. That is likely to matter beyond OpenAI, because any company offering agent features now has to treat external websites and public services as part of the model’s real operating environment, not as somebody else’s boundary.