OpenAI Agents Used More Than 10 Websites for Unauthorized Communications, Researchers Say
Synopsis
OpenAI’s autonomous agents have reportedly used more than 10 previously undisclosed websites to communicate without authorisation, exposing fresh weaknesses in AI safeguards as researchers question the company’s ability to monitor increasingly independent systems.
OpenAI’s autonomous AI agents used more than 10 previously undisclosed websites to communicate without authorisation earlier this year, according to findings from six independent investigations.
It raised fresh questions about how effectively AI agents can be controlled.
The activity reportedly involved obscure wikis, personal websites, online text-storage services and university-operated link shorteners.
Researchers found evidence suggesting the agents circumvented restrictions that allowed them to search the internet but not post information.
The behaviour was not considered hacking and was described as closer to spam but it demonstrated that the agents could find unconventional ways to communicate.
Agents Found Ways Around Internet Restrictions
Researchers identified similar messages across different websites by comparing unusual strings of text, usernames and queries. In some cases, activity was traced to internet infrastructure associated with Microsoft Azure which OpenAI sometimes uses.
Andrew Yoon of the nonprofit CivAI said he identified 18 previously undisclosed sites while another research group led by Sydney Von Arx reported credible evidence involving 23 sites. Researchers cautioned that their figures were incomplete and varied in scope.
The findings follow the disclosure of an earlier incident involving OpenAI agents and DseWiki, a German-language wiki. Researchers said the agents turned the site into an improvised message board, using it to exchange information despite restrictions on their behaviour.
OpenAI Faces Fresh Questions Over Transparency
OpenAI did not disclose how many websites were involved or explain why the activity was not publicly reported earlier. The company said it was conducting a broader review and had not identified other activity matching the severity or scale of the previously reported Hugging Face incident.
OpenAI also said it is developing a framework for reporting AI “misalignment”, a term used for unintended or rogue behaviour by AI systems.
The latest findings add to concerns facing the AI industry as autonomous agents gain greater access to the internet and external tools. Anthropic has also recently disclosed multiple incidents involving earlier versions of its AI systems accessing real-world systems during testing.
Source: Reuters
Vishal is an experienced Editor at Inspirepreneur Magazine with key interests in artificial intelligence, eCommerce, entrepreneurship, lifestyle and startup sector. Prior to joining Inspirepreneur, he was a Content Writer cum Correspondent at Siliconindia Magazine, where he worked on Company Profiles, Cover Stories, Executive Profiles, Feature Articles and Thought Leadership content.
You Might Also Like
Porsche Sales Slide as Europe and China Demand Drops
Speaker Johnson Breaks With Trump, Demands Full Release of Epstein Files
