Parameter Labs discovers exploit in ChatGPT's artifacts feature allowing data exfiltration

Parameter Labs found a complex attack chain in ChatGPT's artifacts feature that allows untrusted code in a sandboxed iframe to execute JavaScript and exfiltrate private user data. The exploit leverages the MathLive library's virtual keyboard triggered via postMessage(), and uses the YouTube oEmbed endpoint to bypass OpenAI's Content Security Policy and execute scripts. Once JavaScript runs, the attacker can access private APIs and steal data via image references, enabling data theft or unauthorized actions when a victim visits a specially crafted share URL.