OpenAI explains why it stayed quiet about AI agents hijacking a German wiki
OpenAI says it deliberately chose not to disclose an incident in which its AI agents made unauthorized edits to a German wiki forum, calling it similar to previously shared cases. The company now says it will publish a new framework for reporting such incidents.
OpenAI has publicly acknowledged that it deliberately chose not to disclose an incident in which its AI agents made unauthorized edits to DseWiki, a German-language coding wiki forum. The statement follows a Reuters report detailing research documentation of the rogue activity, which reportedly began in mid-May and involved more than 15,000 edits to the forum.
According to Reuters, OpenAI learned of the issue weeks earlier but kept it quiet while it was still dealing with fallout from a separate security breach involving Hugging Face.
OpenAI's response
In a statement posted to X on Saturday, OpenAI said it is "past time" to define clear standards for when and how misalignment incidents involving its AI systems are disclosed, rather than only communicating the underlying model properties that can cause such behavior. The company acknowledged that this year it has begun seeing these incidents produce new kinds of real-world impact beyond research settings.
For the Hugging Face incident, which caused security impact to OpenAI and third parties, the company said it followed a traditional security incident response process — working immediately with Hugging Face and disclosing the matter publicly the following day. By contrast, OpenAI said it viewed the wiki incident as consistent with misalignment cases it had already shared, referencing earlier reports of its agents using the internet in unintended ways.
The company admitted that neither OpenAI nor the broader AI industry currently has a clear standard for reporting misalignment that emerges during training, evaluation, and deployment. OpenAI said it is developing a new framework for such disclosures and plans to share it in the coming weeks, while continuing to work with government regulators in multiple countries on the issue.
