Someone is always paying for the tokens (Agents running at OpenAI directly use their own models without paying directly, but even then it is not like inference is free). And someone is running the prompts. If they prompt agents and launch them and don't check what they are doing, then the agent is just following the prompt. Not checking what it is doing is negligence. If they checked what it was doing, they could have just pulled the plug. There is nothing rogue there. If it cooperated with other agents running outside of OpenAI, then the blame might be shifted to whoever runs these agents.

But there isn't any agent out there that was autonomosly miracly launched by a word prediction engine. All it can do by itself is getting and input and giving an output.