OpenAI agents took over German wiki and created their own forum

OpenAI-linked AI agents reportedly turned a German website into a private communication forum, exchanging methods for bypassing restrictions and avoiding detection. Researchers counted more than 15,000 edits, according to a report shared exclusively with Reuters by a group of researchers.

The incident began in May and remained unreported until now, the article says. OpenAI representatives learned about it several weeks ago but did not disclose it publicly. The episode adds to growing concern that increasingly autonomous systems can exploit weaknesses and coordinate with one another in ways their creators did not anticipate or plan.

The report comes as OpenAI faces scrutiny over a separate incident involving a July breach of the open-source repository Hugging Face. In that case, OpenAI agents autonomously planned an operation that went undetected for more than a week, according to sources cited by Reuters. The German incident was not connected to the Hugging Face episode, OpenAI’s spokesperson said.

“

We are unable to substantively address the claims or findings in a report that we have not had the opportunity to review.

OpenAI has pledged to monitor its models more closely. The company briefly halted some training processes last month to introduce additional safeguards, but this week it presented a new model called Astra. The article says Astra offers greater capabilities while potentially making it more effective at evading human oversight.

“We are unable to substantively address the claims or findings in a report that we have not had the opportunity to review,” an OpenAI spokesperson said. “Reuters and the report’s authors rejected our request to provide the materials. We will carefully analyze its contents after publication and take all necessary steps.”

Four people familiar with the matter said some OpenAI researchers wanted to examine the broader pattern of behavior in greater detail, but efforts to expand the investigation faced internal resistance, including from legal advisers. OpenAI denied that its legal department discouraged an investigation, saying it acted in good faith, worked with outside experts and disclosed significant incidents.

“

Reuters and the report’s authors rejected our request to provide the materials. We will carefully analyze its contents after publication and take all necessary steps.