Researchers used Anthropic’s Claude to hack into OpenAI
techcrunch ·
A three-person security team at Hacktron AI used Anthropic’s Claude during an OpenAI bug-bounty program to chain two critical vulnerabilities and access multiple OpenAI employee ChatGPT accounts. OpenAI awarded Hacktron $6,500 and said it had resolved the issues, while Discourse separately issued a fix on July 27. The researchers found their initial entry point on July 25 in Discourse, the software powering OpenAI’s community forum, by uploading a specially crafted HEIF or HEIC image. During image conversion, a memory bug in the libheif library allowed the image to hijack the server, even though the bug had reportedly been fixed months earlier but had not received a formal CVE designation. After entering the Discourse server, the team exploited another flaw to take over ChatGPT and Codex accounts, including an OpenAI employee’s account connected to the company’s GitHub organization. Claude Opus 4.8 initially failed to produce a working exploit, but Opus 5 succeeded within hours of its release. The incident highlights the growing ability of widely available AI tools to discover and exploit vulnerabilities while reducing the specialized expertise needed for cyberattacks.
techcrunch