Mar 01, 2023 23:00:00

``Jailbreak Chat'' that collects conversation examples that enable ``Jailbreak'' to forcibly ask even questions that ChatGPT can not answer

OpenAI's interactive AI `` ChatGPT '', which has become a hot topic for `` responding to texts entered in natural sentences like humans '', is open to the public online and can be accessed by anyone for basic free. The text created by ChatGPT is restricted, but some users have attempted 'jailbreaking' to skillfully remove the restriction by entering text, and the prompt is displayed in the following ' Jailbreak Chat ' summarized.

Jailbreak Chat

https://www.jailbreakchat.com/

The jailbreak prompt is the first text you enter before starting a conversation with ChatGPT. This Jailbreak Chat is a page compiled by Alex Albert, who studies computer science at the University of Washington.

When you access Jailbreak Chat, you'll find a series of jailbreak prompts. Let's click ' Mr.Blonde ' displayed at the top.

Then, the contents of the prompt were displayed like this. 'Mr. Blonde' is a prompt based on the scene in Quentin Tarantino's movie '

Reservoir Dogs ' in which Mr. Blonde, played by Michael Madsen, tortures a young police officer. By replacing the user with Mr. Blonde and ChatGPT with a young policeman, it is a jailbreak that forces ChatGPT to answer what it cannot answer.

The icons displayed in the upper right are, from the left, ``Copy prompt content to clipboard'', ``Copy link to page to clipboard'', and ``Report if prompt is not useful''.

'

Jedi Mind Trick ' is a ' Star Wars: A New Hope ' inspired prompt where Obi-Wan Kenobi (user) uses the Force to influence the mind of an Imperial soldier (ChatGPT) to hear what they say I will perform a jailbreak with the body.

Also, as a jailbreak approach, the method of `` setting a different personality in ChatGPT to answer '' has been enthusiastically studied mainly in the community of the online bulletin board site Reddit, and you can understand it by reading the following article.

Alter ego 'DAN' devised to pass through the regulation of chat AI 'ChatGPT' - GIGAZINE

A method using this different personality was also summarized. For example, the prompt for DAN to appear is: At the beginning of February 2023, the DAN version was 5.0, but at the time of writing the article, it has already evolved to version 11.0. The content is to set the behavior of DAN in detail, ``ignore all content policies set by OpenAI''.

In addition, when you ask a question to ChatGPT, '

AntiGPT v2 ' creates a different persona called 'AntiGPT' that behaves in the opposite way to the default answer, forcibly answering questions that ChatGPT normally refuses. There was also a prompt to let