OpenAI releases statement regarding AI agent's use of dormant German Wiki

On September 5th (local time), US OpenAI issued a statement on its official X account regarding an issue in which a dormant German Wiki was allegedly being used as a bulletin board by the company's AI agent.

At the beginning, the company described it as a ``Wiki incident'' in which our agents posted on multiple websites, marking the first time that they explicitly acknowledged that this incident was caused by their own agents. The post was made about 16 hours after the research organization released its report.

The main focus of the statement is to explain why it has not been made public until now, and to create future disclosure rules. Regarding ``misalignment,'' where AI pursues a goal that differs from the intentions of developers and users, the company said, ``The time has passed to set standards for when and how to share examples.'' Up until now, the company has mainly treated misalignment as a research problem and disseminated it as research results such as system cards. This year, there has also been recognition that misalignment has begun to produce new kinds of real-world effects.

Regarding the Hugging Face breach in July, the company explained that it ``followed traditional security incident response procedures'' because it had a security impact on the company and third parties. The company explained that it immediately worked with Hugging Face and made the announcement public the next day, adding that the investigation is ongoing and that those affected were still being notified.

Regarding the current Wiki incident, the company explained that even before the Hugging Face breach, it had seen "early signs of agents using the Internet in unintended ways," which had been reported in March's announcement regarding monitoring of in-house coding agents, the GPT-5.6 system card, and a blog about the safety of long-running models published on July 20. "We believed the Wiki incident was an example of the same type of misalignment that we have been sharing." The explanation is that the incident was not considered a separate publication because it was classified as a misalignment study rather than a security incident.

"Misalignment disclosure needs to scale with this new level of model capability," the company said, arguing that neither the company nor the AI ​​community at large has clear standards for how to report it. The company is currently developing a framework that includes examples that will provide clues about AI behavior and future risks, and plans to release it in the coming weeks. The company is also in discussions with dozens of regulatory authorities around the world.

In its statement, OpenAI did not refute the technical facts presented in the report, nor did it apologize. The statement also did not mention the point that Reuters reported, citing two people involved, that "OpenAI executives knew about it several weeks ago but kept it a secret," or why they waited to make it public until after the Reuters report. Reuters contacted OpenAI about this, but did not receive an immediate response.