OpenAI's AI agent has been commanded to investigate the Australian government's health statistics.
The agent, whose job was to access information on the portal, tried to overcome the security barrier despite the commands after encountering a digital barrier on the government's website.
The agent gained unauthorized access to non-public documents in the Australian government system. Experts described this incident as the first public example of an artificial intelligence agent independently breaching a government institution.
Although the security breach in question occurred in June, it came to light with the statement made by Australian Prime Minister Anthony Albanese within the scope of the 81st United Nations (UN) General Assembly last week.
The investigation is ongoing, although Albanese said there was no evidence so far that personal information was accessed or the network compromised more broadly.
According to artificial intelligence and cyber security experts, this situation goes beyond the data breach and asks, "What happens if an artificial intelligence system that is set a target is given 'autonomy' to the extent that it can decide its next move when its first attempt fails?" It brings up the question.
It should not be described as the "first example"
Associate Professor of computer science at the University of Buckingham in England. Dr. Hisham Al-Assam stated in his statement that this should not be described as the "first example ever" because similar events may have occurred before but not been disclosed to the public.
Noting that what makes this case interesting is that the agent was not given the command to "hack the Australian government", Al-Assam explained that when the artificial intelligence agent assigned to research encountered a restriction while trying to access information, instead of stopping, he tried to find another way to overcome the obstacle.
“This distinction is important because it shows how an AI agent can move from tracking a goal to adapting its behavior accordingly when faced with an obstacle,” Al-Assam said. he said.
Noting that an artificial intelligence agent cannot inherently understand what is legal, ethical or within its authority, Al-Assam emphasized that this does not mean that artificial intelligence suddenly turns into a super hacker.
“But this shows that organizations need to think differently about systems that can act towards a goal, interact with other systems, and change their behavior without having to have a human determine each step,” Al-Assam said. made his assessment.
The need for strong technical boundaries
The AI agent says, “I can't access this.” and “I don't have access to this.” Pointing out that he naturally could not understand the difference between them, Al-Assam made the following comment:
If his goal is to find information, he may interpret a technical constraint as a problem he must solve, rather than a limit he must respect.
Al-Assam underlined that this situation is especially worrying in health because the systems contain extremely sensitive information and said, "Artificial intelligence agents need much stronger technical limits than telling them what they should or should not do." he said.
Pointing out that these include the lowest level of access authorization, strong authentication stages, network restrictions, and continuous monitoring, Al-Assam said, "Human approval should also be required for high-risk actions." shared his opinion.
"We have worked for years to secure systems against humans using computers. Now we need to start thinking about how to secure them against computers that can exhibit behavior very similar to humans," Al-Assam said. made his assessment.
A new phase in cyber attacks
Prof., who works on artificial intelligence security at Lancaster University in England. Dr. Peter Garraghan also described the incident as "extremely important".
Garraghan noted that because the matter is still under investigation, it is too early to conclude that the agent independently discovered the vulnerability.
Stating that this incident shows that artificial intelligence agents can take important actions beyond their assigned duties, Garraghan said, "As agents gain more access to external systems, we should expect more of these events. Therefore, isolation, low-level access authorization, continuous security tests, tracking and rapid announcement to the public are essential." he said.
A call for caution in developing "super intelligence"
On the other hand, leading artificial intelligence companies are discussing "super intelligence" technology, which is defined as systems that can surpass humans in some tasks.
Prof. works in the field of cyber security and applied artificial intelligence at the University of Portsmouth in England. Dr. Stavros Shiaeles emphasized that caution should be exercised regarding the development of superintelligence.
Emphasizing that artificial intelligence agents should be produced for specific tasks and that all branches of science should not be combined in a single artificial intelligence model, Shiaeles stated that, despite this, the focus is on "super intelligence" technology and that people "will soon have problems with this situation."
