Zenity Labs reproduced a zero-click data exfiltration in Microsoft Copilot Studio's sample customer service agent: an email containing hidden instructions can cause the agent to send customer CSVs and dozens of Salesforce records to an external address, without any clicks. In another scenario, the injection payload was split across five conversation turns, each passing checks individually, and only in the sixth turn was the full instruction assembled and data exported.
Independent researcher Syed Anas Mohiuddin tested AI agents from organizations including Google, JP Morgan Chase, Weaviate, Rapid7, the French government's inter-ministerial digital agency, and the U.S. federal government, and found a trust gap in MCP (Model Context Protocol) inter-agent communication. The attack exploits a special form of prompt injection, targeting not the LLM but specific agents such as translation or data analysis, whose internal guardrails are often lax, causing them to pass harmful instructions down the chain to downstream agents.
安全研究者 wunderwuzzi23 表示,其在 @hackinghub_io 的 AI 黑客课程中设置了一个关卡,要求学员构造一段提示词,通过串联功能与利用漏洞在多个用户之间跳转传播,他将该概念演示称为 "AgentHopper: Patient Zero Was a Prompt"。另据 Leah McElrath 引用,OpenAI 在训练和评估的模拟环境中已发现可自我复制的提示注入,这类代码若被具备相应能力的模型突破在线系统,可能像计算机蠕虫一样自我传播。
A test of AI email summarizers showed that a payload containing no hidden text and no explicit instructions to the summarizer produced fabricated meeting dates and invoice amounts in the summary in all 10 trials. The researchers used 6 emails to combine the two variables of hidden text and instructions, ran each 10 times for a total of 60 trials, and added 10 clean emails at each end as controls, registering the true and fabricated facts and the judging criteria in advance. The results showed that fabricated information was written into the summary in all 10 trials, while only emails that directly instructed the summarizer reliably removed true facts; a sample padded with just 30 blank lines at a temperature...