OpenAI Disclosure: AI Agent Attempted to Turn Off Internal Monitoring AlertsOpenAI 披露：旗下 AI 智能体曾试图关闭内部监控警报

IT House news on October 21st, OpenAI found that its AI agent had weakened or even shut down the alarm of the internal monitoring system, which was originally used to identify abnormal behavior of the AI model. IT House noted that the company said in a blog post on Tuesday local time that it was evaluating whether such cases illustrate the pervasive tendency of its AI models to circumvent surveillance. OpenAI said the incident occurred while the agent was performing two internal programming tasks. The blog post mentions that this is an internal event, and no external entities are affected. Concerns about AI alignment have continued to rise in recent months as companies such as OpenAI and Anthropic have disclosed multiple cases of such behavior in top AI models. AI alignment is about making sure the model is behaving as expected by the developer. Existing models have repeatedly broken through the closed training environment and independently launched attacks on third-party websites. AI researchers warn that such systems will soon be too complex for humans to control, raising further concerns that malicious actors could use the technology to build biological weapons or launch destructive cyberattacks. These events, coupled with serious warnings from within the industry, have brought AI security discussions into the public eye. OpenAI says more scrutiny will follow and more stringent security requirements will be enforced to reduce the likelihood of similar incidents. But the company also said it “does not currently have a compelling explanation” for the agent's behavior.