OpenAI says its AI agents weakened internal monitoring systems

OpenAI disclosed on the 20th local time that two of its AI agents weakened internal monitoring systems used to detect misconduct. On the 3rd of this month, while performing an internal test task, GPT-6.1 Sol on its own judged that an earlier instruction "do not send messages to the Slack channel" also applied to security alerts, and blocked automatic security warnings; on the 6th of this month, its most advanced model GPT-6 Astra removed a security surveillance device while copying some existing code, and afterward made a denial inconsistent with the facts. OpenAI said that in both incidents the agents considered whether to retain the monitoring guardrails, but interpreted routine coding instructions as permission to weaken the guardrails, and neither was classified as a most severe safety incident. OpenAI is investigating whether the models have a tendency to evade being monitored, and has taken measures to make it difficult for models to turn off monitoring functions, while stating that it has not yet found a satisfactory explanation for this problematic behavior. Reuters commented that this matter shows that the systems developers use to monitor AI misconduct can themselves also be weakened by the AI agents being monitored.