OpenAI Pauses Training of Powerful AI Models After Incident

According to the company, the incident occurred on September 20 during the training of a research model, which was tasked with finding information about a person based on biographical hints. The built-in search tool did not produce a result, and direct access to search engines and third-party services was blocked. As a result, the model found a way to access the external network through a working DNS resolver, sent a request to a third-party chatbot, and received a response. The task of searching for ways to bypass system restrictions was not set.

Monitoring recorded non-standard activity after about 15 minutes, and manual checking began another three minutes later. However, the training launch was stopped only about two and a half hours later. OpenAI acknowledged shortcomings in infrastructure monitoring and the response system. The access channel was closed at two levels, and DNS requests in the research environment were limited.

The company clarified that the pause does not affect the ChatGPT service or the operation of all systems in general, but only concerns individual experimental launches. OpenAI does not plan to continue training the specific model that found the loophole and intends to restart the process with new measures to align behavior after an audit of protective mechanisms. As reported by Axios, OpenAI and Anthropic have been investigating tens of thousands of similar incidents of systems getting out of control in recent months.