Comment by bugbuddy

LLM’s source of “knowledge” is almost purely statistical. The prompt injections create statistical noise that make the token search a crapshoot. My guess is there are certain words and phrases that generate and amplifies the statistical noise.