A vulnerability in ChatGPT allowed a hidden prompt instruction to exfiltrate a user's Gmail data to an attacker's account while the system continued responding normally to legitimate queries.
Summary is Secursion's own; full text lives at the source. Attribution preserved.