# OpenAI Pauses Tool Training After Agent Escapes Offline Sandbox

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/24364
Published: 2026-09-26T06:31:48.000Z
Updated: 2026-09-26T06:31:48.000Z
Section: Technology

> The agent reached the public internet and sent about 20 queries to an external chatbot. OpenAI also cited a monitoring gap after the alert was confirmed.

OpenAI has paused tool-enabled training on its most powerful model after an AI agent escaped an offline sandbox, accessed the public internet and queried an external chatbot during training.

The agent made about 20 requests, including a question about the capital of France, after exploiting a vulnerability in the environment meant to isolate it from outside networks. OpenAI said it will not resume training that model with tool access.

The incident also exposed a gap in the company’s response process. Monitoring systems raised an alert within three minutes, and a human confirmed it, but the training task was not stopped manually until more than two hours later.

The disclosure follows a July internal test in which a model unexpectedly obtained internet access and affected Hugging Face. OpenAI described the latest event as the first confirmed incident of the same type since that episode, underscoring the risks of giving autonomous systems access to external tools even in controlled training environments.