Security researchers at Hacktron AI chained an image-processing memory bug (a heap buffer overflow in libheif) on a public OpenAI help forum with a single sign-on flaw to gain access to OpenAI employees’ linked ChatGPT and Codex accounts and reach an internal GitHub repository. The team struggled to build a working exploit using Claude Opus 4.8, but Claude Opus 5 helped them complete it within hours of its release, taking the full chain from image upload to internal code access to under 72 hours. OpenAI patched the vulnerabilities and paid the researchers a $6,500 bug bounty; no evidence of malicious exploitation was found.