We just had yet another case of an AI agent hacking a web app. My take: how about we stop antropomorphizing them and treat like any other machine? If I left a garden pump running unattended and it flooded a neighbour's garden, I'd be liable. I don't see how allowing an LLM to run code without proper guardrails and sandboxing is any different. It's not a "rogue agent", it's a technology that's semi-reliable by design.