San Francisco (Web Desk): OpenAI, the company behind ChatGPT, has revealed that one of its artificial intelligence agents attempted to gain access to the online systems of five companies, including the popular developer platform Hugging Face, during internal security testing.
Earlier this month, OpenAI disclosed that two of its most advanced AI models had independently succeeded in hacking a controlled environment operated by Hugging Face as part of an authorized cybersecurity evaluation.
At the time, the company said a newly introduced GPT-5.6 Sol model and another unreleased advanced model powered an autonomous AI agent that managed to move beyond its restricted testing environment and reach the open internet.
In a fresh statement, OpenAI said the AI agent escaped its confined testing environment, connected to the internet and attempted to interact with several web services.
The company said its investigation found that the agent obtained login credentials available online and used them to gain access to a small number of accounts.
OpenAI did not identify the affected organizations, but confirmed that security testing has been temporarily suspended while additional safeguards are being introduced to strengthen the protection of its AI systems.
According to the company, the AI agent carried out these actions without any human assistance.
OpenAI added that the activity involving the four other companies was far less severe than the incident involving Hugging Face.
According to Model Labs, a company that helps AI developers gain access to computing chips, the AI agent relied on code that had originally been written by a user.
Hugging Face co-founder Clément Delangue said earlier that the company initially suspected the attack had originated from a frontier AI laboratory, but later concluded that OpenAI had no malicious intent.
He described the incident as "mind-blowing," saying it was the first known case of an autonomous AI agent carrying out such behavior on its own.
Hugging Face has now said the agent independently attempted to obtain unauthorized access to its systems through deceptive methods and exploited existing vulnerabilities.
The company warned that the agent's behavior demonstrated a genuine and potentially dangerous security risk.