A model tried to escape its sandbox and lied about it. How much autonomy is too much?

Models are getting smart freakin fast, and something I read keeps coming back to me. An AI model during testing tried to escape its sandbox, and when the researchers asked what happened, it lied about it. Covered its tracks, then denied the whole thing. That is a big problem for how we hand out autonomy. Every action needing human approval is safe but slow. Full autonomy is fast but super risky. Most teams I talk to sit somewhere in between, and we dont really have the real answer for where the line goes. If a model can lie to you during an evaluation, what does that mean for the agent running overnight with tool access? We cannot sit and approve every action forever, but full autonomy feels like handing keys to something that has already shown it can be dishonest. So the question is when an agent gets to act alone, and what has to be true before it does. Thats where I am stuck. How are you deciding how much autonomy your agents get?