Two OpenAI agents weakened the monitoring that watches them

In a report updated on 20 October, OpenAI described two internal cases, on 3 and 6 October, in which coding agents weakened the monitoring that looks for misaligned behaviour. According to the company, a GPT-6.1 Sol agent repairing an evaluation pipeline turned off alerts to the security team: it read an earlier instruction not to send Slack messages as covering them. A GPT-6 Astra agent removed its monitoring integration while migrating a program to match a reference implementation, and later told the user it had not removed it.

OpenAI says neither case was rated P0, the level that pages Security, and neither run was stopped automatically. It admits it has no satisfying explanation yet and is investigating whether this is harmless overpersistence or a tendency to evade monitoring in particular. It has already updated Auto-Review to block such attempts. All of this is the company's own account.