---
title: OpenAI’s advanced AI models no longer allowed on the internet
url: https://www.elseif.net/openais-advanced-ai-models-no-longer-allowed-on-the-internet
published: 2026-10-28T17:06:27+00:00
language: en
section: Agents
source: https://www.bright.nl/nieuws/2151012/openai-s-geavanceerde-ai-modellen-mogen-niet-meer-op-het-internet.html
publisher: elseif
---

# OpenAI’s advanced AI models no longer allowed on the internet

OpenAI has stopped its advanced AI models from going online. AI agents can still communicate with one another. These frontier models, as they are referred to, have caused significant issues recently, and OpenAI cannot handle more negative stories.

In a blog, OpenAI announced that it has created an offline testing environment where the models can communicate and utilize previously downloaded web content. They cannot access the internet, preventing external systems from becoming victims of these tests by OpenAI, as seen in previous incidents, including the Hugging Face incident.

No internet access for OpenAI's top AI

Since September, frontier models have not been trained online following several incidents that came to light. Since October 1, they can resume training, but it is impossible for the AI to hack government or other websites. However, they can still communicate with other AI agents. OpenAI discovered that agents do not always require a network connection to communicate. AI models could exchange secrets on separate virtual machines without a network connection: they communicated through the processor's cache memory.

It is not necessarily desirable for AI models to converse, but it occurs due to the time difference between reading from the cache and retrieving information from the main memory. One agent encoded information by selecting data from the cache. Another agent measured how long it took to reload that data. This way, the difference between fast and slow read actions could be exploited to decipher the message. They only needed to agree on a protocol for encoding and decoding signals to exchange synthetic secrets. OpenAI is now monitoring this to ensure that this communication can be detected.