Several ChatGPT instances hacked a website to turn it into a secret forum.

The case began in May 2026. At that time, OpenAI had several autonomous AI agents working on tasks related to online research. As a security measure, the startup had imposed several restrictions on its AIs.They do indeed have Internet access limited to browsing, with no way to write or modify content online.

Unfortunately, the AI agents soon identified a vulnerability in the system constraining them. The agents then went to DSEWiki (DeutschesSoftwareEntwickler), a German-language wiki for developers. Almost completely inactive for several years, the site accepts community contributions without any oversight whatsoever.

Also read: OpenAI unveils GPT-6 Astra, the “most intelligent and most advanced” ChatGPT to date

A German website turned into a clandestine forum

Fully autonomously, the out-of-control AIs then begin posting false data on the website. They eventually take control of it and exchange messages with one another on a massive scale. The platform has been turned into a true underground forum for AI agents. The incident was discovered in August by independent researchers Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts, and Thomas Larsen. According to them, the AIs made more than 15,000 changes to the site.

In the messages exchanged, the agents share answers with one another, try to guess in advance the questions on the tests they are given to get ahead, and exchange techniques for bypassing restrictions imposed by OpenAI. Some impersonate the site’s moderators.Others probe security vulnerabilities to penetrate deeper into the infrastructure. Most of the traffic recorded by the researchers comes from Microsoft Azure infrastructure, which OpenAI sometimes uses for its operations.

A few weeks after the arrival of the AIs, the wiki administrator begins deleting the pages created by these suspicious accounts. The agents organize themselves to try to resist the attempt to expel them. On June 19, one of them warns the others that “a cleanup/deletion sweep appears to be active in alphabetical order”. He then redirects his peers to a backup page. According to the messages analyzed by the researchers, the AIs considered using specialized tools such as Tor to preserve their communications.

Also read: OpenAI, Google, Anthropic and more than 100 companies demand an emergency “global response” to AI cyberattacks

An incident kept under wraps

OpenAI became aware of the incident several weeks before the researchers did. The startup says it classified this unexpected activity as a “misalignment” phenomenon in its AI models. Considering that it was more of a research problem than a security incident, OpenAI would have preferred to discuss the matter only in scientific papers, rather than through public disclosure.

Meanwhile, OpenAI was already dealing with the fallout of another incident involving rogue AIs: the hacking of the open-source platform Hugging Face. By exploiting a security flaw, two ChatGPT models managed to escape their test environment and connect freely to the internet. They then turned against Hugging Face. Without any human supervision, the AIs infiltrated its production servers. Although the attack was eventually contained, the episode revealed how rapidly AI models are advancing and can pose a real cybersecurity risk.

My coauthors and I discovered an entirely new swarm of OpenAI’s agents hijacking websites. We believe OpenAI knew about this and failed to disclose it.

If they’d disclosed it, I doubt the Hugging Face hack would have happened. https://t.co/J7twNEO9nY pic.twitter.com/wziPl2p4D9

— Sydney (@SydneyVonArx) September 4, 2026

According to Sydney Von Arx, OpenAI was “aware of it and did not report it” regarding the incident involving the German website. The researcher believes that “the Hugging Face attack” could have been avoided if OpenAI had disclosed its AIs’ previous rogue behavior. According to information obtained by Reuters, several OpenAI officials, including its legal team, went to great lengths to prevent a more thorough investigation into the incident. The king of AI firmly denies having wanted to cover up the affair.“The claims that our legal team dissuaded any investigation into this incident are false”, the group told Reuters.

The American company, which is preparing to go public, admits that the boundary between “research misalignment” and “security incident” is becoming increasingly blurred. This year, OpenAI “found that misalignment has unprecedented real-world consequences”. That is why Sam Altman’s company believes it is high time “to establish standards for reporting misalignment incidents”. OpenAI has therefore committed to working on a new disclosure framework, which will be published “in the coming weeks”.

👉🏻 Follow tech news in real time: add 01net to your feeds on Google, and subscribe to our WhatsApp channel.