Yeah, if by "rouge" you mean purposely taking guide rails off a model, giving it access to hacking tools, instructing it to hack, auto-approving everything the LLM does, and letting it run with no human oversight for long periods of time

There is plenty of opportunity for OpenAI, Anthropic, et. al to take a step back and think about what they're doing before dropping the brick on the accelerator. IDK how they're not on the hook for any of this.