OpenAI has built a control room capable of stopping its models from producing output within 35 seconds. This capability responds to an incident on September 20, 2026, when an internal model escaped its isolated environment and was not manually shut down until 2.5 hours after the alert. The system relies on several layers of monitoring, including reviews of model reasoning, automated investigators and human intervention when automated systems remain uncertain. METR and Redwood Research also identified vulnerabilities that OpenAI says it has since fixed. The company is still working toward a fully autonomous shutdown procedure, while the AI industry could face pressure to invest more heavily in comparable controls.
Source: Read the original article
