ChatGPT Vulnerability Allowed JavaScript Execution and Full Account Takeover via Shared Conversation Links
Security researchers at Parameter discovered a critical flaw in ChatGPT that enabled remote code execution in a logged-in user's browser simply by clicking a shared conversation link. The attack chain combined the platform's HTML preview rendering feature with the MathLive library's unvalidated postMessage handling and a YouTube oembed callback to bypass CSP restrictions. Successful exploitation granted attackers access to private conversations, the ability to open chats, and full control over the victim's account. OpenAI acknowledged the issue as P1 severity, deployed a server-side fix within eight days, and paid a $5000 Bug Bounty reward. No user action was required as the patch was applied centrally. The vulnerability stemmed from the interaction of otherwise secure components rather than a single isolated flaw.
ekurakNov 5, 2026Translated from Polish