Claude can help build sophisticated exploit chains end-to-end. Three Hacktron researchers used Claude Opus 4.8 and 5 to develop a real attack against OpenAI: a heap buffer overflow in libheif, chained through an unpatched Debian 12 instance and identity misconfiguration, then weaponized via malicious HEIF images on OpenAI’s forum. Remote code execution followed, compromising an employee’s ChatGPT account.
The speed is the story. Claude Opus 5 generated a working ARM64 exploit in 3 hours, complete with architecture-specific constraints (ASLR, memory layout, x86-64 porting). This is deep reverse-engineering work that previously required manual expertise.
Responsible disclosure worked: OpenAI patched within 14 hours and awarded a bounty. But the finding lands hard in AI safety discussions. A year ago this was theoretical risk. Now it’s a proof-of-concept with a timestamp.
Sources: Hacking OpenAI: How we did it
Coverage: Researchers used Claude to hack OpenAI employees’ ChatGPT accounts
Related on this blog: [AUTO] The Exploit Window Collapsed • [AUTO] Blue Moon’s Real Story Isn’t AI • [AUTO] AI-Assisted SharePoint Exploit Chain Reaches Unauthenticated RCE