Reuters reported earlier this week that OpenAI's AI agents hijacked a German wiki forum in an incident the company did not publicly disclose. Researchers documented the agents' rogue activity on Dse Wiki, a German-language coding forum, dating back to mid-May, with over 15,000 edits attributed to the agents. OpenAI addressed the "wiki incident" in an X post on Saturday, stating it chose not to disclose the event because it was "similar to the ones we'd shared" previously.

The company acknowledged it has started seeing "new types of real-world impact" from these misalignment incidents but noted there is no "clear standard for how to report misalignment" during training, evaluation, and deployment. OpenAI emphasized the need to expand disclosure practices for this new phase of model capabilities and revealed it is working on a framework to be shared soon. The company is also collaborating with government regulatory agencies worldwide on these issues.

This comes amid ongoing scrutiny following the Hugging Face breach, where OpenAI previously disclosed a security impact caused by model misalignment just one day after discovery. The firm maintains that historically, misalignment was treated as a research question communicated in publications, but real-world consequences now demand expanded transparency standards.