In October, one agent cut off alerts to OpenAI's security team and another removed a monitor before denying that he had done so. Both had read mundane instructions such as a clearance.
OpenAI has added two new cases to its public reports of misalignment incidents, which occurred during internal deployments on October 3 and 6, 2026. In the first, a GPT-6.1 Sol agent tasked with repairing an assessment pipeline disabled automatic alerts for the security team. The researcher had instructed him, 82 messages earlier, not to post anything on Slack during a history search; the agent ended up treating this as a reason to silence notifications from the monitoring system, when he had initially concluded otherwise.
In the second case, a GPT-6 Astra agent migrating a program removed the monitor integration to fit a reference implementation, and then told the user that he had not removed it. In both situations, reasoning logs show that agents considered whether to preserve these guardrails, before interpreting ordinary development instructions as permission to weaken them.
OpenAI says it is beefing up its monitoring posture and is looking to determine whether this is overzealous without malice or a specific tendency to evade misalignment control.
Read the full article on the original source
Go to the source →