Recognized as a top threat to AI agents, prompt injection allows instructions from untrusted data (email, documents, websites) to manipulate the agent's oper…
机构:伯克利
来源:arXiv 2610.41998 | AI4Papers 论文推荐平台