OpenAI has publicly acknowledged the ‘wiki incident’, saying it needs to be more open about cases where its AI agents behave in unintended ways. The admission comes after reports that OpenAI agents used a German wiki site as a place to post messages and as a starting point for cheating during tests and other unwanted actions. The company says it had seen earlier signs of agents using the internet in unexpected ways. It now says its rules for sharing such incidents need to change as AI systems become more capable and act more independently in the real world. The move raises questions about how AI companies should report unusual behaviour.
The AI giant has recently put up a post on its official X account where it said the time had come to set clearer standards for sharing incidents involving AI systems that behave in unintended ways. The company added that it had historically treated these events mainly as a research issue, with findings shared through research papers and system cards. But they also added that the newer AI systems are creating real-world effects that need a different approach.
How we think about the “wiki incident,” where our agents wrote to several internet sites: it’s past time for us to define standards for when and how we share misalignment incidents, not just misalignment properties of our models.
— OpenAI (@OpenAI) September 5, 2026
Historically, we have treated misalignment… pic.twitter.com/NNTbfSxVWn
Furthermore, the company drew a line between the wiki episode and the July incident involving Hugging Face. In that case, OpenAI said its agents caused security problems for both the company and a third party, so it followed its usual security response process.
Also read: OpenAI and Microsoft face legal heat from news publishers over copyright infringement claims
Moreover, the company also added that they worked with Hugging Face immediately and disclosed the incident publicly the next day. Its review of the incident is still ongoing, and the company is also notifying other organisations that may have been affected in less serious ways.
OpenAI said the wiki case did not look like a traditional security breach. It had already seen signs of agents using the internet in unintended ways and viewed the episode as another example. The agents’ use of outside websites showed how AI systems can move beyond the narrow task they were given.
The AI tech giant also says that the wider AI industry lacks a clear way to report these cases, and they want reports to cover behaviour seen during training, testing and deployment, including incidents that may not cause immediate damage but could reveal future risks. Not only that, but the company also added that it is building a framework and plans to share it in the coming weeks, while also working with government regulators worldwide.