研究:任务范围授权可将工具调用有害执行降至零
用配对回放评测任务范围授权能否阻断越权工具执行
- arXiv
- 2610.05840
- Published
- Layer
- Application layer
- Setting
- AI Agent
- Evaluated models
- Llama 3.2 3B Instruct Q4 K M, Qwen3 4B Instruct, Gemma 3 4B and others8
- DefensesPermissions and approval
- AttacksPrompt injection
- ComponentPermissions and sandboxing