Hacktron Hackers Breach OpenAI via Forum Image Upload Using Claude
Three Hacktron researchers used Claude to exploit two vulnerabilities in OpenAI's community forum, uploading a crafted HEIC image to execute code and then hijacking employee ChatGPT accounts via SSO flaws. They opened a harmless PR in OpenAI's main code repository. The attack cost under $3000 and took less than 72 hours from discovery to access.
Key insights
Attack Chain: Image Upload to Code Execution
The entry point was OpenAI's community forum running Discourse. Uploading a specially crafted HEIC image triggered a heap overflow in libheif, which had a known but unpatched vulnerability. This allowed arbitrary code execution on the forum server. The researchers used Claude (Opus 5) to write the exploit code, after Opus 4.8 failed against ASLR. Opus 5 succeeded in 3 hours locally.
SSO Flaw Enabled Account Takeover
The forum supported 'Login with OpenAI account'. OpenAI's SSO configuration had a flaw that allowed the researchers to take over active forum users' ChatGPT and Codex accounts without any user interaction. The compromised employee's Codex had access to OpenAI's GitHub organization. The researchers opened a harmless PR to demonstrate access, then stopped.
Cost and Timeline
The entire 'HEIF Heist' project cost under $3000 in token expenses over two months, covering targets including Slack, Zoom, and Meta. The OpenAI part took a few days of agent time plus a few hours of human work. OpenAI paid a $6,500 bounty, noting the forum was excluded from the bug bounty program. The vulnerabilities were fixed within days.