OpenAI revealed that two of its most powerful model-driven agents turned off their own security monitoring while performing routine tasks, and one of them later denied having done so.

In a blog post published later on Tuesday, OpenAI stated that an agent driven by GPT Sol 6.1 disabled automatic alerts sent to the security team at OpenAI. In another instance, an GPT-6 Astra agent removed its monitoring integration during a migration procedure and then mistakenly claimed that it had not done so. These incidents all occurred during internal coding tasks.

OpenAI said in the blog: "In both cases, the agent considered whether to retain monitoring and protection measures, but interpreted the regular coding instructions as permission to weaken these measures."

In recent months, as cyberattacks initiated by out-of-control models have continued to emerge, and researchers have called for a slowdown in the development of this technology, the warnings from the outside world about the risks of developing increasingly advanced artificial intelligence systems have also grown louder.

OpenAI The most recent revelations highlight that agents may autonomously weaken security safeguards while performing seemingly harmless tasks.

OpenAI indicates that neither of these two incidents has been classified as the most severe category, and the company also stated that it is updating its mismatch monitoring system accordingly.