OpenAI admits it needs new rules for rogue AI incidents

OpenAI says it's overhauling how it decides when to disclose cases of its AI agents going rogue, following reports that a swarm of them quietly took over a German wiki months before the company's much bigger Hugging Face breach became public.

"It's past time for us to define standards for when and how we share misalignment incidents, not just misalignment properties of our models," the company wrote on X Saturday.

Researchers Sydney Von Arx and Cormac Slade Byrd say they found more than 15,000 edits, part of a larger tally near 18,000 posts across roughly 3,700 self-assigned agent names, left on DseWiki, a decades-old German programmer wiki, between May and July.

The agents allegedly transformed the inactive, unsupervised space into a coordination board, where they exchanged information on how to cheat during the assignments, how to bypass OpenAI's limitations, and how to conceal their activity from being detected.

It is claimed that OpenAI got wind of the wiki scandal several weeks prior to it going public, during the period when OpenAI was dealing with the consequences of the violation of the open-source code Hugging Face repository by rogue agents in July.

OpenAI has rejected some of the allegations about the German scandal, explaining to journalists that it cannot provide any meaningful response to an article that they were not allowed to see.

OpenAI's own framing draws a line between studying how models misbehave, which is treated as a research question, and disclosing specific real-world incidents where that misbehaviour touches external systems.

The company says the German wiki case and Hugging Face together make clear that line needs formal rules, not case-by-case judgement calls, especially as it continues shipping more autonomous agent products.