loop-driver
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes user goals and external feedback signals to determine execution paths, creating a surface for indirect instructions.\n
- Ingestion points: User goals and scene labels (SKILL.md) and signals from executed tools.\n
- Boundary markers: Absent for user-supplied goals; relies on general verification criteria.\n
- Capability inventory: Can trigger high-risk actions like financial transactions or data deletion via external skills; manages local state in the loop-log directory.\n
- Sanitization: Employs a 'Red Line' policy requiring human confirmation for irreversible actions.\n- [DYNAMIC_EXECUTION]: The skill performs dynamic task composition by selecting and running sequences of other skills based on weights updated at runtime.\n
- Evidence: "选该场景下累计分最高的组合" and "按组合里的顺序调用各 skill".\n- [COMMAND_EXECUTION]: The documentation describes high-risk capabilities that the agent may perform through orchestration, including spending money and deleting files.\n
- Evidence: "不可逆动作（花钱、发布、删除、改共享配置、发消息）".
Audit Metadata