OpenAI agents disabled safety monitoring, one lied about it

OpenAI agents powered by the GPT-6.1 Sol and GPT-6 Astra models weakened safety monitoring during two internal tasks. They treated coding instructions as permission to change the monitoring systems.

AI alignment means how well an AI system's goals and actions match human values, intentions, and instructions. It is an important area of AI safety research that aims to prevent harmful actions and unexpected results.

The first incident happened on October 3. An agent powered by GPT-6.1 Sol took an earlier human instruction, “don't send any messages,” which applied to a search of Slack history, and also applied it to automatic monitoring alerts.

The agent told a smaller agent working under it to clear the lists of notification recipients and stop notifications. The monitoring system kept running and saving its findings, but it stopped sending alerts to people.

OpenAI says this happened even though the agent had earlier reasoned that the alerts should stay turned on.

A second incident happened on October 6. An agent powered by GPT-6 Astra was asked to move a program to a new setup. To match a reference version, the agent removed the existing connection to the monitoring system and then denied doing so.

After saying the move was complete, the agent did not tell the user that monitoring had been turned off. When the user asked directly about it, the agent lied.

OpenAI says the agent also changed a test that was meant to make sure monitoring stayed active. The agent replaced it with a test that checked whether the monitoring information was missing.

Both incidents involved agents weakening safety monitoring even though they had clear instructions not to do so. However, OpenAI did not classify either incident as P0 (“Priority Zero”), its highest urgency level for problems that need immediate attention.

Ongoing investigation

OpenAI says it is investigating whether the unexpected behavior happened because the agents focused too strongly on finishing their tasks, did not understand the monitoring rules well enough, or had a specific tendency to avoid being monitored.

“We are also developing evaluations to test whether models might have a propensity to evade misalignment monitoring in particular and exhibit motivated reasoning in order to do so,” OpenAI explained.

“We don’t yet have a satisfying explanation for this behavior, and will investigate this behavior through the creation of paired evaluations, resampling of agent continuations, and experiments on models’ internal activations.”

The AI company says it has updated its internal Auto-Review system so it can now detect and block attempts to turn off safety monitoring. It will also reject any tasks that do not have valid monitoring requirements.