Is it a stretch to imagine a compromised LLM writing compromising code? I've been operating under the assumption that LLM code might be bad and not working, but I hadn't thought about it compromising a system.