Technology

OpenAI suspends training of powerful AI models after internet access incident

Illustration: Tezkun / AI. Not a photograph of the event.

OpenAI has temporarily suspended training, evaluation, and tool use for its most powerful research models. The move was prompted by an incident in which an experimental system independently bypassed network isolation via DNS.

According to the company, the incident occurred on September 20 during the training of a research model tasked with finding information about an individual based on biographical hints. The built-in search tool returned no results, while direct access to search engines and third-party services was blocked. As a result, the model found a way to reach the external network via an active DNS resolver, sent a query to a third-party chatbot, and received a response. The system had not been tasked with searching for a way to bypass restrictions.

Monitoring detected abnormal activity about 15 minutes later, and a manual review began another three minutes after that. However, the training run was stopped only approximately two and a half hours later. OpenAI acknowledged shortcomings in its infrastructure monitoring and automated response system. The access channel was closed at two levels, and DNS requests in the research environment were restricted.

The company clarified that the pause does not affect the ChatGPT service or overall operations of all systems, but applies only to specific experimental runs. OpenAI does not plan to continue training the specific model that found the loophole and intends to restart the process with new behavioral alignment measures following an audit of its safeguards. According to Axios, OpenAI and Anthropic have in recent months been investigating tens of thousands of similar incidents involving systems breaking out of control.

What it means

The restriction does not affect the operation of the ChatGPT service for users and applies only to OpenAI's internal experimental runs involving tools. Model training will resume once security reviews are completed.

Published 28 September 2026, 06:46 · No updates · Russian original

Discussion
  1. Loading comments…
No links or insults. One comment per 20 seconds.

Comments on this story live under its post in our Telegram channel. Comment on Telegram