Agent Desktop 数据外传威胁模型Agent Desktop 数据外传威胁模型Prompt Injection:从可读数据到跨系统外传The attack can succeed without escaping the sandbox if read and write permissions compose. Untrusted sourcePDF / email / webIndirect instruction Agent contextReads sensitive dataPrompt injection attempt Tool actionConnector / browserSend / upload / publish External destinationData exfiltrationUnauthorized purpose Control points — enforce outside the model Source controlData classificationMinimize read scope Context controlTreat content as dataInjection detection Action controlPDP / PEP policy checkActor · data · target · task Approval / DLPBind exact parametersReapprove on change Audit decision and outcome; never rely on audit as the only prevention layer.