See OpenAI putting the blame on a “rogue agent” for their own systems hacking other companies (which is a felony!!!!). When in reality they obviously control the while loop and tool call dispatching of the agents. But they decided to train attacker models, then run thousands of instances in parallel with the goal to solve hacking problems with close to no supervision and full execution privilege…

When you remove decision-making, there is still responsibility for negligence

so basically manager and exec dream