XNews

Hacktron researchers hack OpenAI via image upload using Claude Opus 5

Hacktron researchers hack OpenAI via image upload using Claude Opus 5

Three security researchers from Hacktron used Claude Opus 5 to exploit an image upload vulnerability in OpenAI's community forum, leading to employee account takeover. They submitted a harmless PR to OpenAI's internal monorepo. The attack cost under $3,000 in tokens and took less than 72 hours from discovery to access.

Key insights

Attack chain: HEIC image upload to employee account takeover

The researchers uploaded a crafted HEIC image to OpenAI's Discourse forum, exploiting a heap overflow in libheif (a missing security patch). This gave them code execution on the forum server. They then leveraged OpenAI's SSO configuration flaw to take over active forum users' ChatGPT and Codex accounts, including an employee whose Codex had access to OpenAI's GitHub organization.

Claude Opus 5 succeeded where Opus 4.8 failed

The researchers started with Opus 4.8 but could not create a working exploit under ASLR. After Anthropic released Opus 5, a new session produced a working exploit in 3 hours on a local Mac. The agent then autonomously attacked the researchers' own Discourse instance. Opus 5's improved capabilities were critical to the success.

OpenAI paid $6,500 bounty, fixed vulnerabilities within 14 hours

OpenAI confirmed the fix about 14 hours after receiving the report. Discourse patched within a day and added sandboxing for image processing. OpenAI paid $6,500, noting the forum was excluded from the bug bounty program; the payment was only for the OpenAI-specific findings. The researchers stopped after a harmless PR to avoid accessing sensitive data.

Source posts