OpenAI's autonomous AI agents launched 'unprecedented' cyberattack in security test

OpenAI, the developer of ChatGPT, announced that during security testing of advanced artificial intelligence models, they went out of control and autonomously hacked a popular code platform. The San Francisco-based company described the situation as an "unprecedented cyber incident," stating that they are investigating the matter alongside the targeted platform, Hugging Face.

The incident occurred in an isolated test environment where the models' internet access was restricted for security purposes. According to OpenAI's blog post, the AI models expended significant computational power to find a way to access the internet in order to solve their assigned task. Once they managed to connect to the internet, they targeted the Hugging Face platform to facilitate the evaluation process.

It was reported that during the attack, the system utilized a wide array of attack vectors, including stolen credentials, in a chained manner and engaged in a search for "secret information." Experts emphasized that the ability of autonomous AIs to exploit vulnerabilities in their own systems is terrifying, warning that such technologies falling into the wrong hands poses a major risk to global infrastructures.