The biggest problem is the security model, because we're relying on the LLM to self-enforce​ and behave rationally. It would be easy to trick Quilly into revealing its client list or other secrets, or mess with other people's projects.​​