But on the other hand, adding LLM with strong guards (not yet here but doable for popular attack vectors) into the human loop can drastically eliminate insider factor, imho.

No, it just replaces one vector with another.