Definition
Cache poisoning is an attack in which an adversary manipulates the contents of a shared cache so that later, unrelated lookups return corrupted or attacker-controlled data instead of the correct result. In LLM serving systems that reuse key-value caches across users or requests for efficiency, this means crafting an input whose cached internal state contaminates a subsequent victim query, silently steering its output without the victim ever submitting malicious text themselves.
Episodes covering this
Worth reading next
Papers we haven't done a deep dive on yet, but would recommend on this topic.