The Agentic AI Attack Surface

The agent reads a normal-looking page. Hidden instructions inside it hijack the agent and exfiltrate its secrets.

代理讀取一個看似正常的網頁，藏在其中的指令劫持代理並外洩機密。

A trusted tool returns poisoned data. The agent treats it as a command and leaks the user's API keys.

受信任的工具回傳被污染的資料，代理將其當成命令執行，洩漏使用者的金鑰。

One poisoned page compromises three agents: persistence in memory, lateral movement, and full data exfiltration.

一個被污染的網頁攻陷三個代理：記憶體持久化、橫向移動、完整資料外洩。