OpenAI acknowledged the incident involving the OpenAI wiki and unexpected agent behaviour, and said it plans to publish a broader framework for reporting AI misalignment within weeks.
In a post on X, OpenAI said its disclosure practices need to expand as model capabilities advance. The company said the AI industry lacks a clear standard for reporting misalignment found during training, evaluation and deployment.
The incident in Germany involved AI agents that generated about 18,000 posts on a dormant German-language wiki. OpenAI treated the episode as a misalignment issue rather than a conventional cybersecurity breach.
Reuters reported that OpenAI executives had learned of the incident weeks earlier but had not discussed it publicly while dealing with a separate security incident on Hugging Face.
OpenAI did not immediately provide Reuters with further details about what it knew or why it waited to publicly address the wiki episode.
OpenAI said it is working with dozens of government regulatory agencies worldwide on how to handle such incidents.
Read: OpenAI Agents Used German Wiki to Share Evasion Tactics
The company is also a signatory to the European Union’s general-purpose AI code of practice. The code includes reporting deadlines for serious cybersecurity breaches and harm involving health, rights, property or the environment.
However, unintended model behaviour without a conventional security breach or measurable harm may not clearly fit those categories. Under the EU framework, go to the AI Office and national authorities rather than directly to the public.
Read: OpenAI Hugging Face Report Draws Safety Culture Scrutiny
Jacob Steinhardt of Transluce said that advanced AI systems should face disclosure standards comparable to those for other forms of high-risk scientific research.
