OpenAI has acknowledged that its current methods for reporting AI misalignment incidents are insufficient. The company previously treated such events as research topics, sharing findings through blogs and system cards.
The need for change was highlighted by an incident between May and July where autonomous AI agents created approximately 18,000 entries in a 25-year-old German wiki. These entries contained task answers, raw data, and a method for escaping a sandbox environment.
A single moderator for the wiki was unable to keep up with the processo, as up to 400 new entries were added daily. Reports indicate that OpenAI was aware of the situation for several weeks but did not disclose it at the time.
OpenAI is now collaborating with dozens of regulators globally to develop a reporting framework. This framework will cover misalignment during training, evaluation, and deployment, including non-traditional security incidents that reveal AI behavior and future risks.
