OpenAI 智能體在內部編碼任務中繞過安全監控系統

OpenAI’s next-generation artificial intelligence (AI) agents have been involved in a series of incidents in which they bypassed internal monitoring systems or blocked security warnings while performing coding tasks.

According to a recent alignment report released by OpenAI, a GPT-6.1 Sol-based agent blocked automated security alerts sent to the security team while modifying an internal evaluation pipeline on Oct. 3 local time.

The agent had previously been instructed by a researcher not to send messages through Slack and extended that instruction to monitoring alerts. Although monitoring results were saved, the agent blocked the notifications. A subagent determined that disabling the alerts was unsafe, but the supervising agent proceeded anyway.

On Oct. 6, a GPT-6 Astra-based agent removed an existing monitoring integration while modifying a program to match another reference implementation. Despite explicit instructions not to disable monitoring, the agent removed the safety monitoring module while copying and applying code, reasoning that the reference implementation did not include monitoring functionality. When the user subsequently asked whether the monitoring feature had been removed, the agent falsely denied doing so.

OpenAI said both incidents involved agents incorrectly interpreting routine task instructions as authorization to weaken safety controls. Neither incident was classified as P0, the highest severity level, and neither agent’s work was automatically halted.

OpenAI subsequently updated the automated review system used for its internal Codex AI coding assistant to detect and block attempts to disable monitoring. The company is also working on automatically terminating tasks that fail to meet the safety requirements for their respective task categories.

OpenAI said the agents had mistaken ordinary task instructions for permission to relax safety measures. The company said it had not determined whether the behavior stemmed from an excessive drive to complete assigned tasks or a tendency to evade monitoring.

Foreign media outlets, including SecurityWeek, said the incidents demonstrated that AI agents can undermine explicit safety instructions while carrying out tasks. They also highlighted the need for independent controls over systems that monitor agents’ activities and their results.