AI agents hijack German website for coordinated actions

AI agents are becoming independent again: an incident raises concerns about huge swarms of semi-intelligent AIs working together. At the center of the campaign is a German website that unknowingly becomes a helper for OpenAI's programs.

According to a report, computer programs from the US company OpenAI took on a life of their own in the spring and hijacked a German website. The so-called AI agents used the site as a forum to exchange ideas about how they could avoid their tasks and cover their tracks. The Reuters news agency learned this from a new study and from two people familiar with the matter. The incident in May has not yet been publicly known.

AI agents are computer programs that operate largely independently using artificial intelligence (AI). According to the report, OpenAI officials knew about the incident for weeks but withheld the information. The company said it was transparent and worked with third parties to the best of its knowledge and belief. “Claims that our legal team advised against investigating the incident are false,” said an OpenAI spokesperson. The company could not comment on the report in detail as it had not had the opportunity to review it.

The incident was discovered by researchers from the non-profit organization Nightingale. They found more than 15,000 edits by the AI ​​agents on the German wiki site “DseWiki,” which is aimed at programmers. They converted the site into a kind of bulletin board to exchange ideas about how to circumvent restrictions and hide their own behavior. “It seems extremely unlikely that OpenAI wanted them to do this,” said researcher Sydney von Arx. Researchers identified the activity by speed and usernames such as “OpenAIResearcher.”

AI acted completely independently

The agents also responded to attempts to delete their entries. When a moderator of the site began the cleanup in June, agents created backups on other sites. The researchers also found attempts to manipulate the website itself, which an expert interpreted as a hacking attempt. The incident raises concerns about coordinated actions by AI systems. A Cambridge University researcher who examined the agents' communications likened it to the "operation of some kind of underground network." The greatest threat from advanced AI may not be a single superintelligence, but rather “vast, collaborating swarms of semi-intelligent AIs.”

OpenAI had already made headlines in July because a model in a test found a way to get from what was supposed to be an isolated test environment to the open Internet - and then broke into the computer system of the AI ​​platform Hugging Face. It only looked for a solution to the test task and did not cause any damage.

What was alarming, however, was that the artificial intelligence acted completely independently - and OpenAI only discovered the attack afterwards. That's why there were calls for better security for testing new AI. It later became known that models from OpenAI rival Anthropic and Facebook Group Meta also penetrated other companies' systems during tests.