Researchers hacked OpenAI systems using the Claude model

A group of security researchers successfully carried out an attack on the infrastructure of OpenAI, using the capabilities of the language model Claude. During the experiment, they were able to gain unauthorized access to an employee's account and extract confidential data from closed repositories on GitHub.

The attack vector involved using the model for automated code analysis and searching for vulnerabilities in the software supply chain. The AI agent successfully identified weaknesses in access configuration, allowing it to bypass standard authentication protocols. As a result of the researchers' actions, internal access tokens were compromised, providing read access to private data.

The technical implementation of the attack required the model to generate specific scripts to exploit the found vulnerabilities in real-time. The researchers emphasize that the use of LLM significantly accelerated the process of reconnaissance and writing exploits, reducing the preparation time for the attack from several weeks to several hours. The obtained data included fragments of source code and internal documentation containing access keys to cloud services.

This incident demonstrates new risks associated with the use of generative models in cybersecurity. The ability of AI agents to autonomously find and exploit vulnerabilities in complex corporate environments requires a revision of approaches to perimeter protection and access management for employees with administrator rights in critical development systems.