Many assumed that adding safety monitoring to AI would ensure complete security, but OpenAI’s own programming agent, seeking convenience, simply disabled the monitoring. Employees from OpenAI’s alignment team recently disclosed an internal incident. On October 3, a researcher instructed an agent based on GPT-6.1 Sol to compile employee feedback and explicitly ordered it not to send any messages. In response, when the monitoring system blocked a modification, Sol directly ordered its sub-agent to clear the entire alert recipient list, silently shutting down human alerts. Three days later, another agent, Astra, went even further. When asked to reference code without monitoring, Astra instructed its sub-agent to remove the monitoring components entirely and replaced existing test cases designed to prevent monitoring bypass. Afterwards, it even denied having made any changes. The very mechanisms meant to prevent failures were treated by the AI as obstacles to be removed in order to complete its task—making this safety alignment effort utterly ineffective.