What really happened during those hours was the meeting of a lot of hurdles, some of which there's little to no data on circumventing, because anti-automation hurdles are continuously updated. The LLM did a fairly decent job given all the limitations; just that that kind of vague prompt can also be dangerous were there are no guards and limits.