Researchers used Claude to breach OpenAI systems
A three-person team from Hacktron AI used Anthropic’s Claude in an OpenAI bug-bounty investigation to chain two critical vulnerabilities. The researchers accessed multiple OpenAI employee ChatGPT accounts and reached company software, including an employee’s Codex account connected to OpenAI’s GitHub organization. OpenAI paid Hacktron $6,500 and said it has fixed the issues.
The initial entry point was a weakness in Discourse, the software behind OpenAI’s community forum. Uploading a specially crafted HEIF or HEIC image triggered a memory bug in libheif, a library used during conversion to JPEG through ImageMagick. The bug had reportedly been fixed months earlier, but was not assigned a CVE identifier, and the vulnerable version remained in use. Discourse released a fix on July 27 after the researchers reported the problem.
Hacktron said the Claude Opus 4.8 version available to its security researchers initially failed to create a working exploit. After Anthropic released Opus 5, the same task was completed within hours. The incident has renewed attention on the growing ability of widely available AI models to automate vulnerability research and exploitation.
Written by Sakhanda Wire AI from reporting by TechCrunch. Facts, quotes and full context are in the original article below.