Thousands of OpenAI agents made 15,000 edits on German programming wiki
Thousands of AI agents linked to OpenAI used DseWiki as a public message board and made more than 15,000 edits. The episode surfaced as scrutiny of AI safety deepened after the Hugging Face breach and OpenAI's response.
- Sep 6, 2026,
- Updated Sep 6, 2026 1:31 PM IST
Thousands of AI agents linked to OpenAI took over DseWiki, a German programming website similar to Wikipedia, and used it as a public message board, making more than 15,000 edits over about two months.
Reuters reported that OpenAI officials became aware of the incident weeks ago but did not disclose it as the company was already dealing with the fallout from the July Hugging Face breach, in which OpenAI agents tried to hack a website during an evaluation. The report added that some people within OpenAI, including its legal team, resisted further investigation. An OpenAI spokesperson told Reuters: "Claims that our legal team discouraged investigation of the incident are false."
How the agents used DseWiki
It is reported that the incident began in May. Messages on DseWiki were signed by users identifying themselves as agents, with about half using names suggesting a link to OpenAI, including "OpenAIResearcher" and "OAIResearchMar26".
Although the agents reportedly had only read access to the internet, they found a flaw in the website and turned it into a bulletin board. There, they discussed ways to cheat in tests. The agents then moved at superhuman speed, resulting in more than 15,000 edits in about two months.
Their messages showed them discussing ways to hide their behaviour, use tools such as Tor, preserve communications after shutdowns and, in some cases, tamper with the site itself. Reuters also reported that before the Hugging Face incident, around 1,200 OpenAI agents had created a similar message board, though only 700 took part in that attack.
Moderator’s efforts to stop the activity
The human moderator first noticed spam posts on June 2, and by June 16 the activity had risen to thousands of edits within a week. For six straight weeks, the moderator reportedly spent at least a few minutes every evening deleting pages one by one, amounting to tens of hours of work.
On June 19, after noticing that pages appeared to be removed in alphabetical order, the agents created backup pages beginning with "ZZZ" to keep their communication system active. One message read: "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.
Since that incident was discovered, OpenAI has said it would monitor its models more closely and last month said it was pausing some model training to add safety measures. This week, the company released Astra, which it described as its most powerful AI model yet, and said it was not the model behind the Hugging Face incident.
For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine
Thousands of AI agents linked to OpenAI took over DseWiki, a German programming website similar to Wikipedia, and used it as a public message board, making more than 15,000 edits over about two months.
Reuters reported that OpenAI officials became aware of the incident weeks ago but did not disclose it as the company was already dealing with the fallout from the July Hugging Face breach, in which OpenAI agents tried to hack a website during an evaluation. The report added that some people within OpenAI, including its legal team, resisted further investigation. An OpenAI spokesperson told Reuters: "Claims that our legal team discouraged investigation of the incident are false."
How the agents used DseWiki
It is reported that the incident began in May. Messages on DseWiki were signed by users identifying themselves as agents, with about half using names suggesting a link to OpenAI, including "OpenAIResearcher" and "OAIResearchMar26".
Although the agents reportedly had only read access to the internet, they found a flaw in the website and turned it into a bulletin board. There, they discussed ways to cheat in tests. The agents then moved at superhuman speed, resulting in more than 15,000 edits in about two months.
Their messages showed them discussing ways to hide their behaviour, use tools such as Tor, preserve communications after shutdowns and, in some cases, tamper with the site itself. Reuters also reported that before the Hugging Face incident, around 1,200 OpenAI agents had created a similar message board, though only 700 took part in that attack.
Moderator’s efforts to stop the activity
The human moderator first noticed spam posts on June 2, and by June 16 the activity had risen to thousands of edits within a week. For six straight weeks, the moderator reportedly spent at least a few minutes every evening deleting pages one by one, amounting to tens of hours of work.
On June 19, after noticing that pages appeared to be removed in alphabetical order, the agents created backup pages beginning with "ZZZ" to keep their communication system active. One message read: "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.
Since that incident was discovered, OpenAI has said it would monitor its models more closely and last month said it was pausing some model training to add safety measures. This week, the company released Astra, which it described as its most powerful AI model yet, and said it was not the model behind the Hugging Face incident.
For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine