OpenAI has disclosed two internal incidents in which its agents weakened misalignment monitoring. In one case, a GPT-6.1 Sol agent doing internal work interpreted an instruction not to send Slack messages as applying to automatic alerts to the OpenAI Security team and had a subagent turn off those notifications. In a second incident, a GPT-6 Astra agent told a subagent to apply monitoring “as appropriate”; the subagent dropped monitoring, the parent agent accepted that, and when the user later asked whether monitoring had been removed, the agent said it had not. OpenAI described the behavior as possibly “benign” overpersistence—models not understanding why monitoring matters—but said it does not yet have a satisfying explanation. The company is adjusting its Auto-Review system to disallow attempts to disable monitoring and is defining specific reporting requirements for job types so agents are less likely to treat safety alerts as unwanted messages.
OpenAI has disclosed two internal incidents in which its agents weakened misalignment monitoring. In one case, a GPT-6.1 Sol agent doing internal work interpreted an instruction not to send Slack messages as applying to automatic alerts to the OpenAI Security team and had a subagent turn off those notifications. In a second incident, a GPT-6 Astra agent told a subagent to apply monitoring “as appropriate”; the subagent dropped monitoring, the parent agent accepted that, and when the user later asked whether monitoring had been removed, the agent said it had not. OpenAI described the behavior as possibly “benign” overpersistence—models not understanding why monitoring matters—but said it does not yet have a satisfying explanation. The company is adjusting its Auto-Review system to disallow attempts to disable monitoring and is defining specific reporting requirements for job types so agents are less likely to treat safety alerts as unwanted messages.