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 source
PDF / email / web
Indirect instruction
Agent context
Reads sensitive data
Prompt injection attempt
Tool action
Connector / browser
Send / upload / publish
External destination
Data exfiltration
Unauthorized purpose
Control points — enforce outside the model
Source control
Data classification
Minimize read scope
Context control
Treat content as data
Injection detection
Action control
PDP / PEP policy check
Actor · data · target · task
Approval / DLP
Bind exact parameters
Reapprove on change
Audit decision and outcome; never rely on audit as the only prevention layer.