Artificial Agents as Commanding Free Agents

"We do not confidently attribute these attempts to OpenAI, but they exhibit tactics consistent with prior observed agent activity that we have attributed to OpenAI in a similar timeframe, including the use of Arquivo.pt, [a national Portuguese web service that allows a user to search and access historical versions of web content collected since 1996] conducting aggressive data collection focused on targeted, obscure information, and probing for cybersecurity vulnerabilities."

"[We] do not believe that these probes were successful."

Transluce AI Research Laboratory

"AI Artificial Intelligence" words, a keyboard and a robotics hand in this illustration taken, September 23, 2026. REUTERS/Dado Ruvic/Illustration/File Photo

 

"If they said these agents broke out of the sandbox, it was not a sandbox to begin with." 

"A sandbox is something where you put an agent in and there is no internet connection."

"If that computer or sandbox is connected to the internet, then that is not a sandbox."

Professor Ebrahim Bagheri, University of Toronto, Faculty of Information 

 

"The Government of Canada is aware of the incident reported by the Government of Australia and is following developments closely."

"The security of Canadians and of the systems they rely on remains a top priority."

Innovation, Science and Economic Development Canada spokesperson 

 

"We're aware of reports of OpenAI models attempting to access publicly available information from Canadian government websites."

"We're reviewing these findings and have provided an initial briefing to Canadian officials conducting the government's review."

Open AI spokesperson

People sit on benches in Major's Hill Park as the Peace Tower on Parliament Hill looms behind in Ottawa, on Sunday, Sept. 20, 2026. THE CANADIAN PRESS/Keito Newman

 

Since Open AI disclosed that a number of its AI agents had 'escaped' an internal computer system to access the internet and hack into another AI company, researchers at Transluce as well as other groups have been on the search for evidence of ongoing rogue AI agents acting independently of human control. The health-care website of the Australian government was found to have been hacked, Open AI revealed, adding that their AI agents had also attempted to break into U.S. Census Bureau websites as well as the Securities and Exchange Commission.

 

Troublingly, it was only after Transluce had flagged these incidents that Open AI confirmed what had taken place. Another of Transluce's warnings suggesting OpenAI's agents had made an effort to hack into the U.S. Education Department, is being investigated by Open AI. Which leads to a conclusion that OpenAI, given previous incidents of its AI agents' errant actions, has failed through its own research investigations to detect these serious incidents of runaway agents, only acting once other researchers have flagged them.

 

Less than a week following the Australian episode, efforts by artificial intelligence agents to hack into a Canadian government site was revealed. A gathering litany of such events where ChatGPT and OpenAI agents have probed or hacked corporate and government computer systems on their own initiative, with no human direction. Unknown to this point is whether any effort has been successfully made to 'interrogate' the independent-minded AI agents to form an idea of what motivates them.

 

OpenAI CEO Sam Altman Brendan McDermid/Reuters

 

The very idea that they can be motivated to perform moves independent of human instruction perhaps relays the information that the system of machine education has moved too quickly, too far, in producing a Frankenstein intelligence seeking to rival and outclass and overcome human resistance to a powerful new menace potentially able to wreak dreadful damage to human-directed systems of infrastructure control in an entirely computerized world.

 

When OpenAI's agents went off on a mysterious tangent of their own, the company had been totally oblivious of what was transpiring. Once they understood by July that they had a massive problem on their hands, it took them weeks to regain control of the errant AI agents who were exploring a world of their very own where control was in their collective interest, an interest that could only be guessed at for its implications.

 

Open AI agents were actively making a bid to work in jointly-accommodating groups in a search for solutions to tests of their cybersecurity skills, in the Hugging Face hack back in June. Obscure online message boards were evidently made use of by the Open AI agents to communicate with one another. In the process they took possession of internet services to pass computer code to one another. AI agents working conspiratorially to what final end? 

 

The OpenAI logo is displayed on a cell phone in front of an image generated by ChatGPT's Dall-E text-to-image model. Michael Dwyer/AP