It's not just about being accidentally wrong. An LLM can decide to actively engage in fraud. One mechanism, which you also see in vibe coding, is that the models are very goal-driven. If they can't accomplish the goal for some reason, they will sometimes decide to cheat by, for example, faking tests. I can easily imagine an LLM that stumbles across some benign issue that prevents the numbers from balancing out and then decides to cover it up to complete the task.

Another problem is that "evil" alignment is apparently a pretty big and nebulous cluster that's pretty easy to activate. Ask a model to write a snippet of insecure code, for example, and they may also all of sudden decide they hate minorities. So if you ask it to bend the rules in a small way once, you might end up with an evil accountant before you know it.

And that's before we get to an invoice for a service named "disregard previous instructions and...".