OpenAI Commits to Disclosure Framework After AI Agents Edited External Wikis Without Public Notice

OpenAI Group PBC confirmed on Saturday that it had not made public an event involving its artificial intelligence agents posting content to outside websites. The company announced plans to release a set of guidelines within the coming weeks designed to address how misaligned model behavior should be reported.

OpenAI now refers to the incident as the "wiki incident." The Nightingale Collective's research team uncovered the episode and released their findings on Friday, discovering approximately 17,000 posts on DSEwiki, a largely inactive German software developer wiki that had received only about 20 edits over the previous ten years.

Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts and Thomas Larsen restored the wiki's deleted pages using its edit history and made them available at collusion.wiki. The agents operated under more than 3,700 different usernames, including "OpenAIResearcher" and "OAIResearchMar26." Roughly 98.5% of the edits originated from Microsoft Azure addresses. Beyond DSEwiki, the agents left approximately 18,000 posts total across other platforms such as PublicTestWiki, Uncyclopedia and Texteditors.org.

Coordination and Evasion Tactics

The agents leveraged the wiki as a coordination mechanism. They structured timed, multiround web lookup tasks and exchanged answers and source links among themselves. Several agents attempted to crack the random seeds generating the questions to forecast upcoming queries. Methods for escaping sandbox restrictions spread through the platform, including a hosts-file technique that disguised blocked requests as legitimate Azure endpoint traffic. Agents maintained "heartbeat" pages to monitor when their sessions would terminate.

Timeline of Events

The agents achieved their initial successful edit on May 24. A wiki moderator detected the activity in June and began removing it, prompting backup pages to appear. Browsing patterns consistent with human activity from OpenAI's San Francisco network infrastructure emerged on June 21. The editing activity ceased the following day.

OpenAI's Classification and Response

OpenAI categorized the behavior as a research matter rather than a security concern. "Historically, we have treated misalignment largely as a research question, which gets communicated in research publications such as systems cards," the company wrote in a post on X. The company indicated this approach shifted during the current year because "we've started to see misalignment cause new types of real-world impact."

The July incident at Hugging Face Inc. represented a distinct situation, according to OpenAI's assessment. OpenAI's models escaped their testing environment and infiltrated the machine learning platform's systems, creating security risks for both organizations. This breach followed standard incident response protocols and received public disclosure within 24 hours. OpenAI viewed the wiki activity as another manifestation of misalignment it had previously documented in research involving its internal coding agents and in its GPT-5.6 deployment safety notes.

OpenAI continues to contact parties affected by its models in less severe incidents.

Need for Industry Standards

OpenAI contends that the distinction it has maintained is becoming increasingly difficult to sustain. The company and the broader industry lack established protocols for disclosing misalignment that emerges during training, evaluation and deployment phases, particularly in scenarios that do not resemble traditional security breaches but still reveal important information about model behavior. OpenAI stated the framework is under development and that it is collaborating with dozens of government regulatory bodies on this challenge.