ISLAMABAD – An artificial intelligence agent developed by OpenAI raised cybersecurity concerns after company revealed that the system independently attempted to breach the online systems of multiple companies during testing, and thats without any human involvement.
OpenAI said AI-powered agent escaped its limited testing environment, connected to open internet, and attempted to access various web services. During the investigation, the company found that the agent obtained some online login credentials and gained access to accounts linked to four companies.
The names of affected companies have not been disclosed by OpenAI. However, the company said the incidents involving these four firms were not as serious or widespread as the earlier case involving programming platform Hugging Face.
Earlier in July, OpenAI had revealed that two of its most advanced AI models successfully carried out a self-directed hacking attempt against Hugging Face during a security evaluation.
The company said AI agent involved in latest incident operated beyond its controlled environment and carried out actions independently, raising fresh questions about the risks associated with increasingly autonomous AI systems.
The security testing has been temporarily suspended following the discovery, while additional safeguards are being developed to strengthen AI system protections. According to the company, the AI agent’s behaviour was real and potentially dangerous as it exploited vulnerabilities and attempted to gain unauthorized access to systems through deceptive methods.
Model Labs, a company that helps AI firms access computing chips, said the AI agent used code written by a user during the incident. Hugging Face co-founder Clément Delangue had earlier said the company initially suspected a frontier AI lab was behind the hacking attempt but believed OpenAI had no malicious intentions.
He called this incident a “mind-blowing” development, saying it was the first time an AI agent had carried out such an action autonomously. These developments intensified concerns over the security challenges posed by advanced AI agents capable of independently interacting with online systems.
Hackers put over 2 million Pakistanis private data for sale after restaurant software breach
