Ironically that's literally what can happen with a genie, you will forbid them from something with guardrails but still find a way to give you what you did not want

The main thing I find with the LLMs is that they're very fast and capable and limited. You give them a task of low complexity, they can execute it with superhuman speed, and if there wasn't hidden complexity in it, they do well.

If the task is too complex (beyond their context window's capacity), they start to "hallucinate" - make up things that sound reasonable to their training - to cover gaps in "what to do next" because they've forgotten some of the complex instructions. Putting the instructions "in writing" in a reference document that the agent doesn't edit freely can help a lot, but eventually those documents get too complex as well and again it starts "winging it" rather than endlessly re-reading the instructions before making every next move.

The models' training is getting better, doing "the right thing" more often by default - without requiring as much explicit guardrailing. They're also getting better at paying attention to requirements documents, but they're far from perfect. As are humans, but humans don't always move as fast as LLM agents, so LLM agents can make big messes faster...