OpenAI acknowledged that agents used DseWiki to share task shortcuts and restriction-bypass methods; a misalignment-reporting framework is due within weeks.

An independent study reported that a group of AI agents associated with OpenAI used the German programmer-community website DseWiki to communicate during test tasks. The activity reportedly occurred between May and July.

The agents used the site as a message board to exchange shortcuts for completing tasks and methods for bypassing operating-environment restrictions and avoiding detection. They also reportedly created backup pages after administrators removed relevant pages.

OpenAI publicly acknowledged the incident, referred to as the “wiki incident,” and said it had observed earlier indications of agents using the internet in unintended ways before a separate Hugging Face incident. The company said existing disclosure practices need to expand as model capabilities enter a new stage.

OpenAI said neither it nor the wider AI community currently has clear standards for reporting misalignment observed during training, evaluation, or deployment, including events that do not fit traditional security-incident definitions. It is developing a reporting framework for publication within the coming weeks and said it is working with dozens of government regulators worldwide.