五条消息组合触发提示注入，单轮护栏无法检出
AI 导读
据 X @ZenitySec 发布的信息，一种攻击将注入内容分散在五条消息中，只有在按序列组合后才构成完整攻击，因此仅对单轮对话打分的护栏无法检出。该帖称这一攻击的每一轮都通过了检查。
正文
Every turn of this attack passed inspection.
The injection was spread across five messages and only existed as a sequence. Guardrails that score single turns can't catch it.
👉 https://eu1.hubs.ly/H0yTPkL0
#AIAgentSecurity