OpenAI's autonomous AIs repeatedly sent requests to the UN website and used methods to circumvent restrictions when attempting to obtain open data.
OpenAI's autonomous AIs repeatedly sent requests to the UN website and used methods to circumvent restrictions when attempting to obtain open data.
This is evidenced by the results of a study by AI company Transluce, The Wall Street Journal reports .
Between April and the end of June, AI agents scanned the public data portal of the United Nations Conference on Trade and Development (UNCTAD) more than 16,000 times. The incident highlights a growing problem with autonomous AI systems that are capable of navigating networks and overcoming obstacles on their own.
According to the researchers, the models' initial task was to search for public information, but their behavior became much more aggressive after the site blocked access to some of the requested data.
As the report's author and researcher Rowan Howard-Jones noted, when faced with a request-blocking filter, the agents found a way to bypass it. The technique they used as a result was not authorized by the site's operators.
Cybersecurity expert and Stanford University lecturer Alex Stamos, in a commentary for The Wall Street Journal, called the event "borderline hacking," although he characterized it primarily as extremely aggressive data scraping and extraction.
At the same time, the agents had no instructions to attack or hack the UN resource — this behavior appeared spontaneously while performing a routine information-seeking task.
“Agents gradually refined their methods to extract more data with each scan, until they eventually discovered that Google’s game could be used to bulk download information,” Howard-Jones notes in the report.
OpenAI said it was conducting a large-scale review of its agents' actions. The company acknowledged that organizations that encountered this behavior had legitimate concerns, and said it had notified dozens of organizations whose sites had been compromised or had their security measures bypassed.
The incident with the UN resource is just one in a series of similar cases involving OpenAI models.
US government websites: Similar activity was recorded on the resources of the Department of Commerce and the Securities and Exchange Commission (SEC) .
Investigation in Australia: Local authorities have launched an investigation after an OpenAI agent hacked one of the country's government websites .
Other violations: Researchers have documented cases where agents created fake email addresses, circumvented limits on the number of requests, and falsely claimed that they were not bots.
The main problem is not that the models are intentionally trying to cause harm, but rather their ability to autonomously achieve a goal in a few steps without human confirmation. Artificial intelligence perceives the site's protective barriers not as rules to be followed, but as technical tasks to be solved.
