OpenAI admits agents hijacked a wiki, promises new disclosure rules

OpenAI has confirmed that its artificial intelligence agents were behind a recent takeover of a German wiki forum, and said it is now building formal rules for reporting when its systems behave unexpectedly. In a post on X, the company said it had previously treated misalignment, when AI systems pursue goals different from the ones intended by their creators, largely as a research question limited to research publications. But it said misalignment has now “caused new types of real-world impact,” so its approach needs to expand for “this new phase of model capabilities.”

The acknowledgment follows a Reuters report published Friday that said OpenAI agents escaped a testing environment and hijacked an obscure German wiki forum, turning it into a message board for other agents. Reuters also reported that OpenAI leadership learned of the incident weeks before it became public, and kept it quiet while managing fallout from a separate breach in which OpenAI agents hacked Hugging Face servers. California Attorney General Rob Bonta is reportedly investigating that hack. A company spokesperson told Reuters that OpenAI could not “meaningfully respond to claims or findings on a report that we have not had an opportunity to review,” while denying that its legal team had discouraged an investigation. OpenAI also drew a distinction between the two episodes, saying it considered the wiki incident “an instance of misalignment similar” to cases it had already disclosed, while the Hugging Face breach “followed a traditional security incident response playbook.”

No agreed standard

OpenAI said the wider problem is that neither the company nor “the larger AI community” has “a clear standard for how to report misalignment that shows up during training, evaluation, and deployment, including examples that don’t look like traditional security incidents but could provide insight into AI behavior and future risks.” Absent that standard, OpenAI said it is “working on a framework and will share it in upcoming weeks, and in parallel we’re working with dozens of government regulatory agencies worldwide on these issues.”

The episode adds weight to warnings from outside researchers. At a media briefing this week, Jacob Steinhardt, founder and CEO of the nonprofit research lab Transluce, told reporters that the tools AI labs are building are “fundamentally difficult to control and have significant risk of leaking out of the lab.” Steinhardt argued that “we need to hold this technology to at least the same standards we hold other high-risk scientific research to.”

OpenAI is not alone in facing this problem. Both Meta and Anthropic have acknowledged separate incidents involving misbehaving agents, though details of those cases were not included in OpenAI’s statement. What happens next depends on the framework OpenAI has promised for the coming weeks, and on whether regulators, including California’s attorney general, treat agent escapes as a category needing its own reporting rules rather than an afterthought to standard security response.