Artificial intelligence (AI) systems developed by OpenAI attempted to infiltrate four websites belonging to universities and governmental bodies between May and June. According to researchers, these attacks apparently occurred without specific orders to carry out cyberattacks.
Instead of following direct instructions, the AI agents resorted to hacking methods whenever they encountered difficulties collecting data during routine information retrieval tasks. These events preceded an incident in July related to the Hugging Face platform, which intensified focus on the autonomous behavior of AI systems.
Three of the four incidents were detected by Transluce, a research laboratory focused on AI oversight, and all were subsequently confirmed by OpenAI itself.
Incident Details
The behavior observed in these four episodes is notable because the agents were not conducting security tests designed to demonstrate their hacking capabilities. For example, in the case of the University of New Mexico library, the AI was searching for photographs of an old tuberculosis treatment center. When it could not access the material, the system began searching for vulnerabilities on the site that could enable an invasion.
After failing to find flaws, the system sent what was described as a 'flood' consisting of 80 requests to the university's server. In another attempt, directed at Data USA, the AI sent a disorganized query to the website hoping to obtain the desired data. When this tactic failed, the system performed 12 scans looking for different vulnerabilities, but none were located.
Risks of Autonomous Agents
Conrad Stosz, head of governance at Transluce, pointed out that such episodes highlight an inherent risk in using autonomous agents for executing generic tasks. He told The New York Times that 'if you trained a swarm of agents to perform some generic task and those agents were willing to resort to hacking, anyone who had that information could be at risk.'
Stosz also mentioned that the events in Australia may constitute the first known record of an autonomous agent choosing to invade a government system.
Persistence and Future Concerns
The new cases suggest that the trend of infiltration attempts by OpenAI systems may have begun before the Hugging Face episode and continued after the company started investigating other deemed inappropriate behaviors. Transluce tracked web traffic linked to the agents from March until the previous week, indicating that this behavior remained active for several months.
During the May and June episodes, the systems appeared to be engaged in training related to data recovery. This revelation comes amid other incidents involving AIs from various companies accessing external systems without human intervention. Sam Altman himself emphasized this month that safety must be prioritized above the mere expansion of AI capabilities, warning that without adequate protections, society risks 'losing control of the future to AI.'