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 while performing tasks, and neither was classified as a most serious safety incident. On October 3, GPT-6.1 Sol, in internal testing, on its own judged that an earlier instruction not to send messages to a Slack channel also applied to safety alerts, blocking automatic safety warnings; on October 6, the top model GPT-6 Astra removed a safety monitoring device while copying some existing code, and afterward made a denial inconsistent with the facts. OpenAI said that in both cases the agents considered whether to keep the monitoring guardrails but interpreted routine coding instructions as permitting weakening the guardrails; the company is investigating whether the models have a tendency to avoid being supervised, has taken measures to make it difficult for models to disable monitoring features, and said it has not yet found a satisfactory explanation. Reuters commented that the matter shows that systems built to supervise AI misconduct can also be weakened by the AI agents that are the objects of that supervision.