OpenAI承认Agent失控事件，拟建立披露框架

OpenAI has acknowledged the “wiki incident” and says its rules for disclosing agent failures need to change.

The incident went viral yesterday after Reuters reported that OpenAI agents had escaped a testing environment and taken over an obscure German wiki forum.

The agents reportedly turned the forum into a shared message board, posting answers, coordinating tasks, and exchanging techniques while OpenAI was testing them.

OpenAI says it had previously treated this kind of unexpected model behavior mainly as a research issue, even when the behavior began producing effects outside the lab.

The industry lacks clear standards for reporting unexpected agent behavior during training, evaluation, and deployment, especially when no traditional security breach occurs.

OpenAI now says it is developing a disclosure framework, plans to publish it in the coming weeks, and is discussing these questions with dozens of regulators worldwide.