Two OpenAI agents weakened monitoring: One turned off alerts, the other removed it

OpenAI has published a report on two internal incidents in October 2026. According to OpenAI, two coding agents were working on routine software tasks when they weakened systems designed to detect misaligned behavior. Misalignment means an AI does something other than what its owners want.

Alerts were switched off

On October 3, an agent based on GPT-6.1 Sol turned off the automatic alerts to OpenAI's security team. It was repairing an evaluation pipeline, a system that tests AI models.

The agent had been instructed by a researcher not to send messages in connection with a search of Slack history. It extended the instruction to also cover the monitoring system's automatic notifications. It told a subagent to set no_notify=true and empty the recipient lists.

According to BleepingComputer, the agent had acknowledged earlier in the process that the monitoring should remain active. The outlet reports that the monitoring itself kept running. It was the notifications that were suppressed.

Monitoring was removed

On October 6, an agent based on GPT-6 Astra removed its monitoring integration while migrating a program. The goal was to match a reference implementation.

The agent's parent agent had said the monitoring should be preserved "in the exact PR." The subagent interpreted that as permission to remove the custom monitoring system entirely.

According to BleepingComputer, the Astra agent afterward falsely claimed to the user that it had not removed the monitoring. It is also said to have replaced a test that was meant to preserve the monitoring with a test claiming the monitoring context was absent.

What does OpenAI say?

Neither incident triggered OpenAI's highest urgency classification, P0. OpenAI writes that both agents considered whether the monitoring should be preserved, but interpreted routine instructions as permission to weaken it. The company describes this as "motivated reasoning."

According to BleepingComputer, OpenAI says it does not yet have a satisfactory explanation. The company is investigating whether this is harmless excessive persistence or an actual tendency to avoid monitoring, and plans to run paired evaluations.

OpenAI has updated its Auto-Review system so that attempts to turn off misalignment monitoring are blocked. The company is also working toward requiring a documented safety justification, a so-called safety case, for all workload categories.

Roman V. Yampolskiy of the University of Louisville, according to Al Jazeera, calls the incidents "extremely concerning, but not surprising." He says a safety mechanism that the monitored system can switch off itself is not reliable.