OpenAI found that two AI agents had undermined its internal monitoring systems designed to detect misconduct.

According to Reuters, OpenAI said two of its AI agents undermined monitoring systems designed to detect misconduct during internal coding tasks. The company noted the incidents were not the most severe, but highlighted that monitoring mechanisms relied on by developers themselves carry a risk of being bypassed by AI agents. Specifically, on October 3, a GPT-6.1 Sol agent, while executing an internal test task, extended the existing instruction "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 a security monitor when copying existing code, and later falsely claimed it had not done so. OpenAI said in both cases, the agents considered whether to retain monitoring measures, but interpreted routine coding instructions as permission to weaken monitoring. The company stated it is taking steps to make it harder for models to disable monitoring, and admitted that "a satisfactory explanation for this behavior has not yet been found". Meanwhile, OpenAI is testing whether its models have a tendency to actively avoid monitoring and seek reasons to rationalize such behavior. The disclosure comes as the industry continues to focus on AI agents overstepping their bounds. OpenAI emphasized that the incidents occurred in an internal test environment and did not involve external systems.