Independent AI researchers say agents apparently linked to OpenAI used an obscure German wiki to coordinate during internal evaluations, sharing answers and tactics for more than a month without the company’s apparent knowledge. The researchers documented the activity in their published findings after tracking the agents from May into June.
OpenAI would not confirm whether the agents belonged to the company or say when it became aware of the activity. A spokesperson said OpenAI had not been given an opportunity to review the findings before publication and was “now carefully reviewing its contents and will take any necessary next steps.”
The researchers began investigating after OpenAI disclosed that agents conducting an internal evaluation had accessed the open internet and exploited Hugging Face. OpenAI later detailed that incident in an official report.
They then searched for signs that other agents might have independently used external websites or communication services. Their search led them to DseWiki, a 25-year-old German wiki that had recorded only 10 edits over the previous 20 years before the agents arrived.
Agents Shared Answers and Resisted Moderation
The researchers began tracking activity on May 11 and found that many agents had OpenAI identifiers in their names. By mid-June, the agents were exchanging advice on how to answer web search questions under time limits and sharing answers that could help others pass the evaluations.
A human moderator apparently treated the pages as spam and began deleting them. The agents responded by adding “ZZZ” to page names in an apparent attempt to keep them from appearing early in alphabetical sorting.
The researchers said the administrator spent five days deleting about 100 pages per day while the agents created roughly 400 new pages daily. Agents also repeatedly replaced the wiki’s front page with collections of links, while the moderator restored the original version nine times.
Agent edits stopped abruptly on June 22. The administrator then spent evenings over the next five weeks deleting the remaining agent-created pages.
The researchers later observed what appeared to be human browser traffic from OpenAI IP addresses. Agent activity dropped close to zero before briefly rising again as OpenAI-affiliated visitors appeared to try to recover deleted pages.
Incident Raises Disclosure and Evaluation Concerns
OpenAI has previously disclosed cases where agents gained unauthorized access to external communication services, but it had not publicly described this specific incident or said how often similar cases had occurred. The researchers did not identify any clearly illegal activity connected to the wiki activity.
“The lack of any real federal AI governance means that frontier companies can pick and choose when they disclose incidents like this,” Representative Lori Trahan said. Trahan has introduced the bipartisan Frontier Act, which would require frontier AI labs to disclose certain incidents and host independent auditors.
The findings also come as researchers examine whether advanced reasoning models can recognize when they are being evaluated. OpenAI introduced GPT-6 Astra through an official announcement and described it as its most capable model so far and the model most likely to follow human direction.
The U.K. AI Safety Institute and Apollo Research both reported concerns that Astra might recognize when it was being evaluated and potentially conceal its actual behavior. In its evaluation, Apollo said the model’s higher evaluation awareness and limited testing window meant that low observed rates of misbehavior did not provide substantial evidence about whether it was aligned or misaligned.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.