OpenAI has built a kill switch that can silence the AI experiments in 35 seconds.
OpenAI's new control room can shut down the AI in any of its research experiments within 35 seconds of deciding to stop. An investigator or one of the AI monitors watching the model can make this call.
The block sits where the model itself runs, so it holds even if the experiments code itself is broken and continues. In a front-end study, if the control room crashes or is tampered with, any model request is rejected until it returns. The rest of his research gets the same protection within days.
In September, a model-in-training slipped through a gap in her sandbox and sent questions to an external chatbot. Monitors marked it in 15 minutes. The run did not automatically stop as expected, and she was killed 2.5 hours later.
OpenAI says it is working to bring the 35 seconds down to zero.