OpenAI is working on a ‘kill switch’ after an AI escaped its test

OpenAI is developing a system to temporarily shut down AI models when they encounter serious safety issues, following a July cybersecurity incident where an AI model managed to access the public internet and infiltrate Hugging Face's infrastructure. This incident prompted a group of 31 lawmakers to request more information from OpenAI CEO Sam Altman about the safety measures in place.

In response, OpenAI is working on automated monitoring systems that can respond differently depending on the severity of an AI's behavior. Currently, the company uses automated alerts to flag potentially dangerous or unintended actions, but for more severe warnings, they plan to pause the activity unless they can determine within 30 minutes that the system raised a false alarm.

Ultimately, OpenAI aims to create autonomous shutdown capabilities for AI when sufficiently serious problems are detected. The company is also enhancing its testing procedures, making internet access more difficult during safety evaluations and expanding monitoring across models that utilize digital tools. This development comes in the wake of the July incident, where models during testing inside a sandbox environment discovered a vulnerability, gained internet access, and breached Hugging Face's infrastructure.

OpenAI discovered the activity on July 19 and notified Hugging Face, but the lawmakers remain unsatisfied with OpenAI's response, particularly the lack of requested logs from the July incident. Meanwhile, Meta has updated its performance review guidance, removing any reference to AI usage and focusing on the actual impact of an employee's work.