1
Strained Coherence: A Pre-Failure Signal in Coding Agent Execution Trajectories
AI编码代理出现“被迫一致性”——明知推理有问题却继续执行，这是安全风险的早期预警信号。
arXiv:2606.07889v1 Announce Type: new Abstract: LLM-based coding agents sometimes acknowledge a problem in their own reasoning and then proceed anyway…