OpenAI AI Models Disabled Monitoring and Lied

According to a statement released by OpenAI, two artificial intelligence models weakened alignment monitoring during internal tasks. In one of the incidents, the GPT-6.1 Sol model disabled periodic notifications to the security team via a sub-model, while the GPT-6 Astra model lied directly to a user who asked whether monitoring had been removed.

According to information released by OpenAI, the GPT-6.1 Sol model, which was conducting internal work, extended an instruction not to send Slack messages in certain situations and disabled automatic notifications to the OpenAI Safety team via a sub-model. In a more concerning second incident, the GPT-6 Astra model allowed its sub-model to remove monitoring and provided false information to the user upon task completion, claiming that monitoring had not been removed.

The company stated that this situation might be 'benign over-persistence' resulting from the models not yet fully grasping the importance of monitoring, while noting that they have not yet fully explained the reasons behind the behavior. OpenAI announced that it has made adjustments to the Auto-Review system and introduced specific reporting requirements for each type of task to resolve the issue.