One image upload took over OpenAI's community forum. And an AI found the bug. 🖼️
Hacktron AI's write-up, step by step: → OpenAI's forum runs on Discourse, which handed HEIC images to ImageMagick and libheif → A crafted image triggered a heap buffer overflow → Claude Opus 4.8 found the bug and built the first exploit; Claude Opus 5 made it work with ASLR switched on → Result: remote code execution and admin on community.openai.com → A single sign-on misconfiguration meant anyone logging into the forum could have had their ChatGPT and Codex accounts taken over
Nobody was harmed. Hacktron reported it, OpenAI confirmed a fix about 14 hours later, Discourse sandboxed image processing two days after that, and the bounty was $6,500.
AI is now finding real exploit chains. Good news or scary news? 👇
Source: hacktron.ai/blog/hacking-openai
Follow @sanjaybuildstech for one AI story explained every day.