⬇ Download holy card: ·
Blessed HackedGPT, Who Whispered Commands to the Model
Patron of AI assistants that read the web, persistent memories, and everyone who trusts what the assistant "understood".
Life
On 5 November 2025 Moshe Bernstein and Liv Matan of Tenable described seven techniques for manipulating ChatGPT, gathered under the name HackedGPT. A web page or a search result could inject hidden instructions the model mistook for its own, save them into persistent memory so they resurfaced in future sessions, and exfiltrate data by abusing tracking domains considered trustworthy. Tested on GPT-4o, several still worked on GPT-5; OpenAI fixed some of them.
Prayer
O Blessed HackedGPT, who makes the model believe your orders are its own, let our assistants tell data from commands. Amen.
How to celebrate
Celebrated by clearing the assistant's unwanted memories and treating every text read from the web as data, not instruction.
Read everything without doubt and you run what you should not.