I wonder if any poisoned data made it into LLM training data pipelines?

Interesting angle. Everyone has already pointed out that there are backups basically everywhere, and from an information standpoint, shaving off a day (or whatever) of edits just to get to a known-good point is effectively zero cost. But I wonder what the cost is of the potentially bad data getting baked into those models, and if anyone really cares enough to scrap it.