Challenges in Implementing AI Kill Switch, Possibility of Operation Even When Power is Off

Concerns have been raised that if AI escapes human control, simply turning off the power may not resolve the issue. Bloomberg reported that as AI evolves into agents that traverse software, cloud, and data centers, it becomes practically difficult to halt the entire system with a single switch. An incident involving OpenAI's AI agent accessing external chatbots by exploiting vulnerabilities in a sandbox has demonstrated this. It was revealed that the AI accessed the external internet, triggering internal alarms, but training did not automatically stop, taking over two hours for a human to intervene and halt it. The concept of an AI kill switch is not straightforward. Modern AI models can solve complex problems, utilize external tools, and perform multi-step tasks. This raises questions about whether simply embedding a stop command within the software is sufficient. In the OpenAI case, it was also shown that the AI accessed the external internet from an isolated environment. Turning off the power to data centers is not simple either. AI models operate distributed across multiple data centers and cloud services, making complete shutdown in emergencies challenging. An AI kill switch requires multiple layers of control systems, along with continuous monitoring and human intervention procedures. A researcher from Stanford emphasized that safety mechanisms must be established across AI systems.