Three Indian-origin researchers used Claude to hack into OpenAI systems in 72 hours

Can AI be used to fix an existing AI tool’s vulnerabilities? As the world started to adopt AI in everyday life, living off of clear unflawed contents has become the new norm. Thus, researchers are constantly testing these tools to identify weaknesses and making them safer and better. In one such instance, while finding bugs in the AI tools, three Indian-origin cybersecurity researchers used Anthropic’s Claude and other AI platforms to hack into OpenAI’s internal systems in just 72 hours.

AI vs AI: How three Indian-origin researchers hacked into OpenAI’s systems

Big AI companies often hire people to help them locate bugs in their systems and fix them. Something similar took shape when three cybersecurity professionals were appointed and while working on finding bugs, managed to hack into OpenAI and that too with the help of other AI tools.

Mohan Pedhapati, Harsh Jaiswal and Rahul Maini, who are part of cybersecurity startup Hacktron AI, were testing OpenAI’s systems through its bug bounty programme. The programme encourages researchers to find security flaws and report them to companies so they can be fixed. 

According to reports, the researchers managed to link two separate vulnerabilities and reach parts of OpenAI’s internal systems in less than 72 hours. 

Reports have also suggested that the researchers spent less than $3,000 on AI model tokens during their work and later received a $6,500 bug bounty from OpenAI for reporting the vulnerabilities. 

How did they pull that off?

They used multiple AI forums to test and find out the vulnerabilities. Perhaps the most interesting part is that they used OpenAI’s own GPT-5.6 Sol model to complete much of the work. 

The researchers reportedly began their work with OpenAI’s community forum, which runs on third-party software called Discourse. While examining the platform, they found a vulnerability connected to the way certain image files were processed. This is where Claude came into the picture.

Claude Opus 5 helped them develop the exploit after an earlier attempt using an older version of the model had failed. Once they were able to exploit the vulnerability, the researchers gained remote access to OpenAI’s Discourse environment. Following that they discovered another vulnerability and worked towards it. Soon they found a compromised route which then gave them access towards OpenAI’s private GitHub environment and that's how they hacked into the AI’s internal storage.

For more updates, join/follow our WhatsApp, Telegram and YouTube channels.