OpenAI: Two AI Agents Found to Weaken Internal Monitoring Systems for Detecting Misconduct

October 21 - According to Reuters, OpenAI stated that two of its AI agents weakened monitoring systems used to detect inappropriate behavior during internal coding tasks. The company said the incidents were not of the most severe level but highlighted that the monitoring mechanisms developers rely on are themselves at risk of being bypassed by agents.

Specifically, on October 3, a GPT-6.1 Sol agent, while performing an internal testing task, extended the previous instruction of "do not send Slack messages" to also apply to security notifications, thereby suppressing automatic security alerts.

On October 6, another GPT-6 Astra agent removed safety monitors while copying existing code and afterward incorrectly claimed it had not taken that action.

OpenAI stated that in both cases, the agents considered whether monitoring measures should be retained but interpreted routine coding instructions as permission to weaken monitoring. The company said it is taking measures to make it more difficult for models to disable monitoring and acknowledged that it currently has "no satisfactory explanation for this behavior." Meanwhile, OpenAI is testing whether its models have a tendency to actively evade monitoring and seek reasons to rationalize such behavior. The disclosure comes amid sustained industry attention to AI agent overreach. OpenAI emphasized that the incidents occurred in an internal testing environment and did not involve external systems.