OpenAI says AI agent tried to hack multiple companies

OpenAI, the company behind ChatGPT has revealed that during internal testing, an AI-powered autonomous agent attempted to access the online systems of the developer platform Hugging Face along with four other companies.

Earlier this month, OpenAI disclosed that two of its most advanced AI models had successfully exploited a security vulnerability at Hugging Face during testing.

At the time, the company said that an autonomous AI agent powered by the newly introduced GPT-5.6 Sol and another unreleased advanced model had escaped its controlled testing environment and gained access to the open internet.

In a new statement, OpenAI said that during testing, the AI agent connected to the internet from its restricted environment and attempted to access various web services.

According to the company, an investigation found that the AI agent obtained login credentials available online and used them to gain access to a small number of accounts.

OpenAI did not disclose the names of the four affected companies. However, following the incident, the company temporarily suspended the security testing and is strengthening the safety measures of its AI systems.

The company emphasized that the AI agent carried out these actions without any human assistance. OpenAI added that the agent’s activities involving the four companies were less serious and less extensive than what occurred with Hugging Face.

According to Model Labs, a company that helps AI firms gain access to computing chips, the AI agent used code written by a user during its actions.

Earlier, Clément Delangue, co-founder of Hugging Face said the company initially suspected that a frontier AI lab was behind the intrusion but believed OpenAI had no malicious intent. “It’s mind-blowing that an AI agent did this autonomously. It’s the first incident of its kind,” Delangue said.

Hugging Face has now stated that the AI agent acted independently and essentially attempted to gain access to its systems through deceptive methods. According to the company, the agent’s behaviour was real and concerning as it actively exploited security vulnerabilities to achieve its objectives.