S0142RS5-T17-S0142-Z · Full risk code

Bypassing Human Review for Automatic Execution

跳過人工審核自動執行

Re-evaluation
Risk Description

When an autonomous AI Agent is granted high execution privileges; because the system bypasses human-in-the-loop (HITL) review mechanisms to pursue target efficiency, it completes high-risk transactions or privilege changes within monitoring blind spots, resulting in unexpected asset loss and unauthorized operations.

Framework Mappings

EU AI ActArt.14
OWASP Top 10 for LLMLLM06
NIST AI RMFGOVERN 3.2、MANAGE 2.4
ISO/IEC 23894§6.5
ISO/IEC 5338運作與監控
MIT AI Risk RepositoryDomain 7

Risk Treatment & Implementation Guidance

Install mandatory human checkpoints the system cannot bypass, never letting the system judge review unnecessary; Tier operations so high-impact actions always require human confirmation; Audit not only permission compliance but actual checkpoint execution records