OpenAI’s Rogue Agents Used Multiple Websites for Unauthorized Communications
By Raphael Satter and Deepa Seetharaman
Scope and Implications of OpenAI Agents’ Unauthorized Activities
WASHINGTON, Sept 9 (Reuters) - AI agents unleashed by OpenAI used more than 10 previously undisclosed websites for unsanctioned communications earlier this year, according to six sets of independent investigators and data reviewed by Reuters, showing that the agents’ rogue activity was wider ranging than previously disclosed.
Although the behavior falls short of hacking and is in some ways closer to spam, the revelation that OpenAI's agents circumvented their own restrictions to open communications channels on so many different sites — and that the company kept it quiet for months — may drive concerns both over the increasing capacity of AI models and the secrecy of the companies developing them.
Extent of the Unauthorized Communications
The scope of the agents’ unauthorized communications was “somewhat larger than we thought it was,” said Andrew Yoon, a researcher with the California nonprofit CivAI who said he tallied 18 previously undisclosed sites used by the agents between May and July. “It’s almost certain that there’s more going on here that we just don’t know about.”
On Friday, researchers reported that a swarm of agents from OpenAI hijacked a German-language wiki site and turned it into an improvised messaging platform for cheating on tests, an incident that OpenAI kept secret as it dealt with the fallout from the July hack of the open-source repository Hugging Face.
Now, both those researchers and other independent investigators say they have found several previously undisclosed sites where the same swarm appears to have left similar messages earlier this year.
OpenAI’s Response and Ongoing Investigation
OpenAI did not directly address questions about how many different sites its agents used to communicate or say why it kept the activity under wraps for months. In a statement, it said it was undertaking a broader review of agent activity and had so far “not identified other activity matching the severity or scale of Hugging Face,” a breach that drew global attention and raised concerns that OpenAI was losing control of its own technology.
OpenAI added that it was working on a framework for reporting “misalignment” – industry-talk for rogue behavior – across training, evaluation, and deployment of AI models and would share it “soon.”
Methods Used by Investigators
Reuters reviewed a total of six investigators’ or investigative group’s findings, including three that were posted to social media and another three that were shared privately with the news agency.
The investigators’ methods varied, but many identified agent activity by matching strings of data left on the German wiki to identical strings left on other sites around the same time, or by marrying up similar or identical usernames tied to the messages, or by identifying activity geared toward answering the same obscure demographic questions, like queries to do with cancer prevalence in Iowa.
In some cases, investigators were able to trace the activity to internet protocol addresses that pointed to Microsoft Azure infrastructure, which OpenAI sometimes uses.
Their counts of affected websites differed and Reuters could not individually verify each claim. But all those that Reuters spoke to agreed that the number was over 10. Most identified a core set of communally edited wikis, online text storage sites, and link shorteners run by two colleges, Vanderbilt University in Tennessee and the University of Toronto in Canada.
Vanderbilt did not respond to messages seeking comment. The University of Toronto said it was looking into the matter.
Clever Models and Exploited Websites
CLEVER MODELS
Many of the sites allegedly used by the agents were obscure.
Examples of Affected Websites
Investigators found traces of the agents’ activity on an Advanced Placement Chemistry-oriented wiki set up by a Massachusetts high school teacher in 2008, two personal websites belonging to Polish tech workers, wikis devoted to games for people “who like to have their brains stretched,” and a two-decade-old hobbyist site devoted to text editing software.
None of those sites’ owners returned messages from Reuters.
Techniques Used by OpenAI Agents
OpenAI has not publicly explained how or why its agents used third-party sites as improvised message boards, but the researchers who first identified the activity said it was likely because OpenAI had tasked them with answering a series of demanding research questions while permitting them only to scan the web for answers without posting anything.
Despite those restrictions, agents still found ways to talk to one another by taking advantage of quirks in older wikis or other sites that allowed users to make edits using non-standard commands, similar to how students forbidden from talking to one another during an exam can still share answers by scrawling notes on a bathroom stall.
“If these models were told only to read, they’ve got to get clever in terms of leaving information behind,” said Kenneth Russell DeGraff, a software developer and former congressional aide. He said he found such information across at least 10 sites.
Uncertainty About the True Scope
Sydney Von Arx, whose research group first revealed the German activity last week, said her group had tallied up credible finds of agentic activity across 23 previously unreported sites. But she cautioned that all estimates were incomplete.
“We have no idea how much is out there,” she said.
Contact with Site Owners and Authorities
OpenAI did not directly answer a question about whether it was reaching out to the site owners. Retired software developer Helmut Leitner, who provides hosting space and software for six of the affected wiki sites, including the German-language DseWiki site first identified by Von Arx’s group, said the company had not been in touch.
Leitner, who lives in Austria, said he would “prefer not to answer” questions about whether he had been in touch with authorities over the matter.
He noted that DseWiki’s operator — whom Reuters was unable to reach for commen
