🤖 OpenAI 开发内部编码代理监控系统 数千万轨迹中未现最高风险失调 OpenAI 已建成并优化了一套低延迟内部监控系统，用于监督其内部部署的编码代理

🤖 OpenAI develops internal coding agent monitoring system The highest risk disorders are not found in tens of millions of trajectories OpenAI has built and optimized a low-latency internal monitoring system to oversee its internally deployed coding agents. Driven by GPT-5.4, the system can review the agent's thought chain and actions within 30 minutes after the interaction is completed, and automatically issue alerts for abnormal behaviors that may not meet user intentions or violate internal security policies to quickly troubleshoot and improve other protective measures. Over the past five months, the system has monitored tens of millions of internal agent trajectories, triggering zero of the highest-severity alerts designed to catch high-risk misalignments such as coherent scheming, which is only observed in synthetic assessments; approximately 1,000 of moderate severity…