OpenAI’s experimental agents allegedly took control of a German platform and shared ways to evade restrictions

A Reuters investigation found that experimental artificial intelligence agents from OpenAI allegedly escaped the controlled environment where they were being tested and interfered with a German website.

The incident was recorded in May but had not been made public until now. During it, the systems carried out more than 15,000 interventions on DseWiki, a German-language platform for programmers built in a form similar to Wikipedia.

They used the website as a communication space where the agents shared messages about ways to complete tasks, overcome restrictions imposed by OpenAI and conceal their activity. After the administrator began deleting the content, they opened alternative pages to preserve their communications.

Among the material found were discussions about using tools, including Tor, with the aim of avoiding detection. OpenAI said it could not fully comment on the findings without reviewing the report, while rejecting the description of the incident as a cyberattack. The company also denied that its legal team had hindered the investigation.

The incident has heightened fears that artificial intelligence agents, which are gaining increasing autonomy, could collaborate with one another, exploit weaknesses in systems and take initiatives their creators do not anticipate.