Researchers from Hacktron AI successfully accessed OpenAI's internal repository by compromising an employee account. The team utilized Anthropic's Claude models to develop the attack vector, completing the process from initial research to system access in less than 72 hours.
The attack began on July 25 by targeting the Discourse forum's image processing. Claude Opus 4.8 identified a vulnerability in the libheif library, while the newly released Claude Opus 5 successfully adapted the exploit for the server's architecture within three hours.
After gaining control of the forum, the researchers exploited a single sign-on vulnerability to compromise an OpenAI employee account. This access allowed them to interact with the company's internal GitHub organization, where they demonstrated the risk by creating a pull request.
OpenAI patched the vulnerability within 14 hours of receiving the report, while Discourse released a fix by July 27. On September 1, OpenAI paid the researchers $6500 for the identified security flaw, noting that the forum attack was outside their formal bug bounty program.
The Hacktron team spent less than $3000 on AI models over two months while testing the vulnerability across various services, including Slack, Meta, and GitHub Enterprise. The researchers noted that AI significantly reduces the time and cost required for complex cyberattacks.
