Security researchers from the startup Hacktron AI successfully breached OpenAI systems using the Claude Opus 5 model. The team participated in the OpenAI bug-bounty program and received a reward of $6,500 for identifying and reporting the vulnerabilities.
The attack began on July 25 through a flaw in the third-party software Discourse, which powers the OpenAI community forum. A memory bug in the libheif library, used for image processing, allowed the researchers to hijack the server after uploading a specially crafted HEIF image file.
The researchers initially struggled to create a working exploit using the Claude Opus 4.8 model. However, after the release of Claude Opus 5, the AI successfully generated the necessary code within hours. The vulnerability in libheif had been previously patched by developers but lacked a CVE identifier, causing it to remain unpatched in the Discourse software.
After gaining access to the Discourse server, the team exploited a second vulnerability to compromise ChatGPT and Codex accounts, including those belonging to OpenAI employees. This access provided a pathway into the company's internal software and GitHub organization.
OpenAI confirmed that the issues have been resolved following the disclosure on July 27. This incident highlights the growing ability of AI models to assist in developing cyber exploits, reducing the time required for such tasks from months to days.
