I don't see this kind of behavior myself, I wonder if it's me not pushing the boundaries enough or me explicitly or implicitly framing the prompt in a way that makes it less likely for the agent to intervene.