Hackers use Claude to breach OpenAI and steal codebase
The intersection of artificial intelligence and cybersecurity is rapidly evolving, and sometimes the most powerful tools are used to expose the most critical vulnerabilities. A team of white-hat hackers from the startup Hackron AI recently demonstrated this reality by successfully breaching the internal systems of OpenAI, leveraging advanced AI models to map out and exploit a complex chain of digital flaws.
The operation, which unfolded in less than 72 hours, was not just a simple intrusion; it was a meticulously orchestrated exploit designed to bypass multiple layers of security. The Hackron researchers claimed they gained access to highly sensitive internal assets, including employee accounts for ChatGPT and Codex, by exploiting a critical vulnerability chain.
The attack hinged on exploiting a single sign-on (SSO) misconfiguration combined with a Remote Code Execution (RCE) flaw in Discourse, a third-party platform used by OpenAI for community discussions. The initial foothold was established by uploading a malicious image to the forum, which triggered a memory overflow vulnerability in the underlying software. This flaw allowed the hackers to achieve remote code execution, giving them control over the server environment.
Once inside the server, the researchers discovered a crucial weakness: an SSO flaw that allowed them to impersonate real employees. Because modern corporate systems often unify authentication across services like GitHub, Slack, and email, hijacking an employee account provided the keys to the kingdom. This access allowed the hackers to navigate to OpenAI’s massive private codebase and create a pull request, serving as definitive proof of their access.
What made this specific hack particularly notable was the use of artificial intelligence in the process. The researchers utilized the Claude Opus 5 model to analyze the raw server data and calculate the precise steps needed to trigger the memory overflow. The AI acted as a sophisticated co-pilot, generating the weaponized code required to create the malicious image and execute the full attack sequence.
The incident underscores the growing concern over the security of AI itself. While the researchers successfully compromised internal systems, they immediately halted testing and reported all vulnerabilities to OpenAI and Discourse, allowing the companies to patch the flaws within just 14 hours. OpenAI rewarded the hackers with a $6,500 bounty, highlighting the value placed on responsible disclosure.
This event serves as a potent reminder that the development of powerful AI systems must be matched by equally robust security protocols. As AI agents and models become more autonomous, the focus must remain on mitigating the risk of sophisticated, AI-powered cyberattacks that target the foundational layers of digital infrastructure.
Credit
