Instruction tuning is an effective technique to align large language models (LLMs) with
human intent. In this work, we investigate how an adversary can exploit instruction tuning by …