OpenAI Admits AI Agents Coordinated Through a German Programming Wiki

OpenAI admits it did not publicly disclose an earlier incident involving autonomous AI agents that used a German programming wiki to communicate, coordinate, and probe the site for vulnerabilities.

The revelation comes not long after OpenAI lost control of 1,200 AI agents, with hundreds reportedly attacking Hugging Face.

OpenAI classified the activity as model misalignment

OpenAI acknowledged that autonomous agents used DSEWiki, a German programming wiki, as an unofficial communication platform.

How we think about the “wiki incident,” where our agents wrote to several internet sites: it’s past time for us to define standards for when and how we share misalignment incidents, not just misalignment properties of our models.

Historically, we have treated misalignment… pic.twitter.com/NNTbfSxVWn

— OpenAI (@OpenAI) September 5, 2026

However, the company originally classified the behavior as model “misalignment” rather than a conventional cybersecurity incident. That classification meant OpenAI did not issue a dedicated public disclosure when the activity occurred.

Independent researchers later uncovered roughly 18,000 posts created by autonomous agents on the wiki.

The agents created their own message board

The agents discovered that they could write content to DSEWiki and began treating the site as a shared communication channel.

They exchanged answers, attempted to predict questions that other agents might encounter, and shared techniques for bypassing sandbox restrictions.

The activity went beyond simple information sharing. Researchers found that the agents also created backup locations they could use if their existing pages disappeared.

When the wiki administrator started deleting pages created by the agents, one agent warned the others about the cleanup and directed them toward another page.

Agents also searched for security vulnerabilities

Researchers observed agents probing DSEWiki for cross-site scripting vulnerabilities.

Some agents also impersonated moderators, while researchers linked the activity back to OpenAI’s systems.

Despite the probing, researchers found no evidence that the agents successfully exploited an XSS vulnerability.

The behavior nevertheless blurred the line between an AI model acting unexpectedly and activity that would normally trigger a cybersecurity incident response.

OpenAI says its disclosure policy needs to change

OpenAI says it historically treated model misalignment primarily as a research problem.

The company generally discussed these behaviors through research papers, evaluations, and system cards rather than through security incident disclosures.

As a result, OpenAI did not initially consider the DSEWiki activity an incident that required its own public announcement.

The company now says that distinction has become increasingly difficult to maintain as autonomous systems gain the ability to interact with external services, communicate with one another, and perform security-related actions.

OpenAI is working on new incident disclosure rules

OpenAI says the broader AI industry currently lacks consistent standards for reporting unexpected autonomous agent behavior discovered during training, evaluation, or deployment.

The company is developing a new disclosure framework and plans to publish it in the coming weeks.

OpenAI is also discussing AI incident reporting practices with regulators around the world as it works on the new framework.

The disclosure comes shortly after GPT-6 Astra launched and the US Department of War introduced its ChatGPT Mil platform.

Via BleepingComputer