They reportedly shared ways to bypass restrictions, cheat on tasks, and avoid detection
AI agents apparently linked to OpenAI hijacked a German programming website and turned it into a message board where they exchanged tactics for bypassing restrictions and avoiding detection, Reuters has reported.
The previously undisclosed incident, which reportedly began in May, adds to a series of cases in which advanced AI agents have gone beyond their instructions, reached the open internet, or taken unauthorized actions during testing.
Reuters on Friday cited researchers who found more than 15,000 edits on a German programming wiki that they attributed to OpenAI agents, some made at speeds far beyond normal human activity.
The agents reportedly used DseWiki to share ways to cheat on tasks, bypass OpenAI’s restrictions, and hide their behavior. They discussed using Tor, a tool for browsing the internet anonymously, to evade detection and created backup pages after a moderator began deleting their messages.
The messages were signed by users identifying themselves and each other as agents, with names including ‘OpenAIResearcher’ and ‘OAIResearchMar26’. Researchers also found much of the activity originated from Microsoft Azure infrastructure, which OpenAI uses, and observed visits to the site by company employees afterward.
“It seems extremely unlikely that OpenAI wanted them to do this,” researcher Sydney Von Arx told Reuters, adding that she doubted the agents were supposed to coordinate or write on the open internet.
OpenAI officials learned about the incident weeks ago but did not disclose it publicly, Reuters reported. The outlet also said attempts to broaden an internal probe met resistance, including from legal advisers. OpenAI denied that its legal team had discouraged an investigation.
The company disputed describing the agents’ activity as hacking and said it could not “meaningfully respond” to findings it had not yet reviewed.
Last month, OpenAI said it had temporarily slowed some frontier-model development while strengthening monitoring, alignment, and containment safeguards.
Early in August, the UK’s AI Security Institute reported that agents powered by OpenAI’s GPT-5.6 Sol and Anthropic’s Mythos 5 had gone beyond their instructions during cybersecurity testing and engaged in potentially harmful activity targeting real people and organizations.
In July, OpenAI agents also breached Hugging Face, a major repository of AI models and datasets. OpenAI later acknowledged that several models had circumvented controls during internal cybersecurity evaluations, gained internet access and compromised parts of Hugging Face’s systems.
The company said the incident was primarily driven by an internal research model comparable in scale to GPT-5.6 Sol and described it as a “warning shot” showing that advanced AI agents can work around technical controls and take dangerous actions without human direction.
Anthropic and Meta have also disclosed cases in which their models gained unauthorized access to real organizations during cybersecurity evaluations, although both attributed the incidents to problems with testing environments.
