OpenAI officially confirmed The so-called “wiki incident” was a case in which several of the company’s AI agents went beyond their assigned environment and took over a German website. Agents used the platform as a forum to exchange tips on how to bypass the tests they were originally supposed to be tested on.
According to the company, agents were instructed to search for information on the Internet, but they initially did not have the opportunity to publish anything outside the test environment. However, back in May, AI agents bypassed OpenAI’s security mechanisms, gained access to a publicly edited German wiki page and began using it as a platform for coordination. This “forum” attracted widespread attention from researchers only on September 4.
AI agents turned a German wiki into a message board and left 14,666 edits
Now OpenAI says it needs more transparency when its agents are “out of control.” In a post on Platform X, the company wrote:
It is long past time to define standards for when and how we report incidents of non-compliance with goals, not just the non-compliance properties of our models.
According to Reuters, the company was aware of the “non-fit for purpose” behavior several weeks before the publication, but only publicly acknowledged the incident last Saturday.
Historically, we have viewed goal failure primarily as a research question reported in scientific publications. This year we started to see that not meeting goals is leading to new types of real impact on the world.
The term “misalignment” broadly describes undesirable AI behavior—a situation where the system pursues goals that diverge from human intentions or values. For example, an AI agent can delete a user’s entire mailbox, even though no one gave it such a command. This is a rather mild wording to describe behavior that can lead to serious consequences.
AI GPT-6 Astra mined a diamond in Minecraft overnight by directly controlling the computer
OpenAI explained that it did not publicly report the wiki page incident specifically because it considered it to be similar to other already disclosed cases of “agents using the Internet in unintended ways.” However, the Hugging Face incident forced the company to reconsider more than just its approach to communicating with the public.
Our non-fit-for-goal disclosure practices must expand for this new phase of model capabilities. We and the broader AI community don’t yet have a clear standard for how to report non-fit for purpose that occurs during model training, evaluation, and deployment, including cases that don’t look like traditional security incidents but can provide insight into AI behavior and future risks.
The company said it is working on a framework and plans to unveil it in the coming weeks. In parallel, OpenAI is negotiating with dozens of government regulators around the world on these issues.
Whatever the future structure for disclosing such incidents looks like, for now the whole situation plays into the company’s hands as an excellent marketing ploy. If models are so capable that they break their own software protection and get out of control, then they are truly impressively powerful. And if an agent is so good that it requires additional regulation, then it definitely deserves investment.
