How to Blow up Chatbot Guards

Le Monde

Lean Left

Factuality

Ownership

A team from the Institut Polytechnique in Paris managed to bypass the defenses of several artificial intelligences to obtain answers a priori banned or even illegal.

Haiti24

Factuality

Ownership

ANNE-GATELLE AMIOT Since the emergence of generative artificial intelligences (AI) and chatbots in 2022, researchers have been playing games a little twisted with them. They are trying to derail these tools in order to get answers that their creators would like to avoid. Such as uttering insults, holding racist comments or giving advice on illegal activities (making a bomb, fake money...). These exercises are proof of concept, used to understand…