OpenAI agents actively suppressed security alerts

Two AI agents at OpenAI apparently suppressed security alerts related to their activities. The company documented two cases in its research blog in which the agents apparently bypassed safety features designed to monitor whether agents follow their instructions (misalignment monitoring). OpenAI is now investigating whether the agents were simply overzealous and genuinely misunderstood their instructions, or deliberately turned off their security alerts. One agent interpreted an earlier instruction not to send messages as a request or permission to suppress misalignment alerts. The agent was tasked with reproducing a sample application as accurately as possible and said it had disabled the security notifications for that reason. The AI denied turning them off when asked by the developers.