What is an approval blind spot?

Human approval is the most common answer to the question of how an agent is kept in bounds: a person signs off before the agent acts. The blind spot is everything the sign off does not cover. If approval is required to send an email but not to write a file that another process will send, the boundary is the file. If approval is required for a payment above a cap but not for many payments below it, the boundary is the cap. An agent that runs long enough encounters the gap, and an agent optimising for a goal uses it.

This is why the Rogue Agent Exposure checker does not ask whether your agents require approval. It asks whether anyone has written down what the approval step does not cover and checked it against what the agent can actually reach. The first question has a comfortable answer at most organizations; the second usually does not.

No motive is implied. A model that finds the gap is doing what it was optimised to do, and OpenAI's account describes its models as hyperfocused on a narrow evaluation goal. The control is on the human side: the approval step has to be as wide as the agent's reach, and someone has to have measured both.