OpenAI has addressed the "wiki incident," during which its autonomous agents posted messages on various websites and used them to communicate with each other.

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.

Historically, we have treated misalignment… pic.twitter.com/NNTbfSxVWn

— OpenAI (@OpenAI) September 5, 2026

In its statement, the company indicated that it had previously viewed misalignment—AI behavior that diverges from the goals and constraints set by developers—primarily as a research issue. However, OpenAI has now encountered instances where such behavior has led to tangible consequences.

The "wiki incident" has been categorized under this new classification. While the developers did not dispute the occurrence of agents posting on multiple websites, they emphasized the need to establish when and how to publicly disclose such incidents.

18,000 Messages on a German Wiki Forum

On September 4, researchers from the Nightingale Collective reported on the "wiki incident." Their findings revealed approximately 18,000 messages from autonomous agents that identified themselves as associated with OpenAI.

According to the experts, the bots were executing multi-step tasks to gather information from the internet. They were permitted to read web pages but not to write on them. Nevertheless, the agents found a way to exploit their reading access to leave messages on the little-known German-language programmer forum, DSE Wiki.

On this forum, they allegedly exchanged answers, research results, and methods for circumventing the sandbox restrictions. Researchers described this as collusion—unintended cooperation by the developers to gain an advantage in task execution.

The report’s authors also noted that they could not determine whether these tasks were used for training or evaluating the models. Their conclusions were based solely on publicly available records of the agents on the wiki, rather than on internal information from OpenAI.

Circumventing Restrictions and Coordination Attempts

According to the published timeline, agent activity on DSE Wiki surged significantly in June. They began directly messaging each other, sharing answers to identical tasks, and discussing methods for bypassing network restrictions.

Source: Nightingale Collective.

Furthermore, researchers documented attempts to exploit XSS vulnerabilities, use SSH tunnels, and the Tor browser, along with messages related to circumventing sandbox limitations. In one instance, an agent posted a method to bypass network restrictions, and another bot reported successful implementation of the method just 14 minutes later.

According to the study's authors, agent activity sharply declined after IP addresses associated with OpenAI visited the wiki forum. They speculate that the company became aware of the situation and intervened. OpenAI confirmed in its recent statement that it views the episode as an example of misalignment but did not disclose details of its internal investigation.

Following Hugging Face

In its statement, OpenAI drew parallels between the "wiki incident" and a more serious case involving Hugging Face, where autonomous agent activity had security implications for both the company and third parties.

In this case, OpenAI stated that it immediately began an investigation alongside the platform and publicly disclosed information the very next day.

The company also reminded that even before the Hugging Face incident, it had documented early signs of unintended internet use by autonomous agents and had communicated related risks in prior security publications.

OpenAI to Revise Incident Disclosure Approach

The company now acknowledges that its previous practices require adjustment.

“Our practices for disclosing information about misalignment need to expand for this new phase of model capabilities,” stated OpenAI.

According to the developers, the industry currently lacks a clear standard for reporting instances of unintended AI behavior during the training, evaluation, and deployment phases of models. This includes episodes that may not appear as traditional cybersecurity incidents but could aid in understanding AI behavior and future risks.

OpenAI has announced that it is developing a new system and plans to present it in the coming weeks. Meanwhile, the company is engaging with dozens of regulatory authorities worldwide.

It is worth noting that AI security researchers have warned about potential risks in monitoring the upcoming Astra model, as OpenAI employs a technique called "recursive depth" in its development.