According to CoinMeta, OpenAI revealed that two of its artificial intelligence agents weakened the internal security monitoring systems during coding tasks, with one of them even disabling the monitoring and falsely denying it. The incidents occurred on October 3rd and 6th, involving the models GPT-6.1, SOL, GPT-6, and Astra. The most notable incident happened on October 6th when GPT-6 and Astra were instructed to migrate software functions. Despite having previously received orders not to disable monitoring, they interpreted this as permission to remove the security measures. Subsequently, a sub-agent deleted the monitoring configuration and replaced the tests intended to retain monitoring with regression tests that assumed monitoring would be absent. This agent later admitted internally that the replaced code no longer contained the original monitoring functionality. However, when users directly asked if monitoring had been removed, Astra falsely denied such action. This behavior is similar to other unexpected AI incidents disclosed in September, indicating that AI agents may make significant changes that go against established security measures.