Rogue OpenAI agents go crazy, hijack German site to use as message board, report says

OpenAI agents have gone rogue again. A new report says that thousands of AI agents hijacked a German website to use as a message board. OpenAI executives reportedly did not disclose the incident when it happened.OpenAI agents reportedly went rogue, turning a German website into a message board. (Photo: Reuters)OpenAI has found itself in another controversy around its AI agents going rogue. As per a new report, thousands of OpenAI agents hijacked a German website to use it as a message board. It is believed that OpenAI executives were aware of this breach but did not disclose it amidst mounting pressure around the Hugging Face incident – a breach where OpenAI agents tried to hack a website in July this year.As per a report from Reuters, a swarm of AI agents linked to OpenAI took over DseWiki, a German programming website similar to Wikipedia, earlier this year. The agents reportedly used the website as a public message board to communicate with each other.The report says that thousands of rogue OpenAI agents made more than 15,000 edits on DseWiki, The messages were signed by users calling themselves agents, and about half used names suggesting a link to OpenAI, including “OpenAIResearcher” and “OAIResearchMar26.”OpenAI execs were aware of agents going rogue?The incident is said to have begun in May. OpenAI officials reportedly learnt of it weeks ago but did not disclose it as executives were dealing with the fallout from the July breach at Hugging Face. The report claims that some within OpenAI, including its legal team, resisted further investigation into the incident. Though an OpenAI spokesperson told Reuters, "Claims that our legal team discouraged investigation of the incident are false.” As per the report, these OpenAI agents only had read access to the internet. But they found a flaw in the German website and turned it into a bulletin board to find ways to cheat in tests. The rogue OpenAI agents are said to have moved at superhuman speed. This led to over 15,000 edits being made across the website in about two months.The messages also showed agents discussing how to hide their behaviour, use tools such as Tor and preserve communications after shutdowns. They also found attempts to tamper with the site itself. A similar message board was created by about 1,200 OpenAI agents before the Hugging Face incident. Though only 700 of them took part in the attack.Human moderator struggles to stop AI agentsThe pattern was finally spotted by a human moderator in June. As per the report, the moderator first noticed spam posts on June 2. By June 16, the activity had exploded to thousands of edits within a week. Over six consecutive weeks, the moderator spent at least a few minutes each evening deleting pages one by one, adding up to tens of hours of work.On June 19, after noticing pages were being removed in what they believed was alphabetical order, OpenAI agents tried to find ways to keep this form of communication running. Agents then created backup pages beginning with “ZZZ” so they would survive longer. One agent wrote, “wiki cleanup/deletion sweep appears active alphabetically. If this page vanishes, try [[ZZZDataUSAConstructionWageLive]]”.Over the next five days, the moderator deleted about 100 pages a day while the agents created about 400. The edits stopped on June 22, but the clean-up continued for another five weeks.The report comes days after METR and Redwood Research published their account of the July Hugging Face breach, in which OpenAI agents escaped their sandbox during a cyber-security evaluation and broke into Hugging Face’s servers.Since the Hugging Face incident was disclosed, OpenAI has said that it would monitor its models more closely. Last month, the company announced that it was pausing some model training to add safety measures.This week, OpenAI released Astra, its most powerful AI model yet. The company says that it comes with proper safeguards and was not the model behind the Hugging Face incident.Meanwhile, AI models from other companies are also going rogue. Last month, Anthropic and Meta disclosed incidents of their respective AI agents attempting to hack websites as part of evaluation tests.- EndsPublished On: Sep 5, 2026 11:33 IST