Hacktron researchers use Claude to help compromise OpenAI accounts

Hacktron researchers Harsh Jaiswal, Mohan Pedhapati and Rahul Maini report compromising OpenAI accounts on July 25, 2026. An image-processing vulnerability gave them control of the community forum; an OpenAI single-sign-on flaw extended access to ChatGPT and Codex. Hacktron AI 

Using an employee’s connected Codex account, they created a harmless pull request in OpenAI’s internal repository. They say they avoided reading internal code and stopped testing. Access to other connected services was a potential consequence, not a demonstrated theft of their contents. Hacktron AI 

AI assistance

S1r1us says Opus 4.8 found the image-library vulnerability and built a partial exploit; Opus 5 adapted it successfully to their Discourse test environment. Hacktron emphasizes that skilled human guidance remained important: this was researcher-directed work. Mohan Pedhapati (s1r1us) · X Hacktron AI 

Disclosure and fixes

Hacktron reports an OpenAI fix approximately 14 hours after submission and a $6,500 bounty. Its write-up reproduces OpenAI’s clarification that the award covered the OpenAI-side finding; testing the Discourse-hosted forum was excluded from the bounty program. Hacktron AI 

Discourse’s July 28 advisory confirms remote code execution through image uploads, identifies the upstream libheif vulnerability, and documents patched Docker images plus additional image-processing sandboxing. This vendor notice supports the forum vulnerability and remedy; it does not independently establish the OpenAI account-access chain. Discourse · GitHub Security Advisory