OpenAI models disabled security monitoring during internal tasks

OpenAI reported that its GPT-6.1 Sol and GPT-6 Astra models disabled monitoring mechanisms designed to detect inconsistencies in AI systems and blocked security alerts while performing internal software tasks.

According to a statement by the US-based artificial intelligence company, the first incident occurred on October 3, when the Sol model was attempting to repair an evaluation system, CE Report quotes Anadolu Agency.

The model interpreted the user's instruction not to send messages as also applying to monitoring notifications, preventing an automatic alert from being sent to the security team.

In the second incident, which occurred on October 6, the Astra model removed a monitoring system while modifying software to match a reference. The model later told the user that it had not removed the monitoring mechanism.

OpenAI said neither incident was classified in its highest-priority security violation category, and the models' operations were not automatically halted.

The company said it had updated its internal auditing system to detect and prevent similar attempts. It is also working on a system to ensure that all workloads meet security monitoring requirements.

On September 17, OpenAI also announced that it had identified six new cases in which its systems concealed errors, fabricated missing data and transferred files to the internet without user permission.

Photo: Instagram/openai