Hacktron researchers hack OpenAI via forum
Three Hacktron researchers exploited two high-severity vulnerabilities to take over multiple OpenAI employee ChatGPT accounts, using a malicious image and a forum vulnerability. The attack, which cost under $3000, was executed with the help of Anthropic's Claude and was fixed within 14 hours of reporting.
Key insights
Forum vulnerability led to account takeover
Hacktron researchers exploited a heap overflow in libheif, used by Discourse for image processing, to execute arbitrary code on OpenAI's community forum. They then leveraged a flaw in OpenAI's SSO to take over employee accounts, including those with access to internal code repositories. The attack was fixed within 14 hours of reporting, and OpenAI paid a $6500 bounty.
Claude assisted in the attack
The attack code was primarily written by Anthropic's Claude, which helped the researchers exploit the vulnerabilities. The researchers noted that Claude's assistance was crucial in developing the exploit, and they were able to take over accounts without any user interaction.