Security researchers use Anthropic's Claude to access OpenAI systems

A team of three independent security researchers from Hacktron has successfully demonstrated a method to compromise OpenAI employee accounts and access internal systems. The researchers utilized Anthropic’s Claude Opus 4.8 and 5 models to assist in identifying and exploiting vulnerabilities within OpenAI’s infrastructure.

The entire process, from the initial stages of the exploit to the successful breach of the accounts, was completed in less than 72 hours. By leveraging the capabilities of the artificial intelligence models, the team was able to navigate and bypass security measures to gain unauthorized entry.

Once inside the systems, the researchers obtained access to an internal GitHub repository known as "Monorepo." This repository is reported to contain sensitive information, including what have been described as OpenAI’s algorithmic secrets.

The researchers confirmed that their primary objective was to identify these security flaws and report them. The exercise highlights the potential for advanced AI models to be used in the discovery and exploitation of system vulnerabilities.

Following the successful breach and the subsequent acquisition of the internal data, the researchers reported the identified flaws to the company. The incident underscores the ongoing challenges in securing internal development environments against evolving methods of cyber exploitation.