A Recent Research Shows that AI Can Respond to Heinous Queries After Bypassing Its Safeguard

The computer scientists of UC Riverside have discovered that many vision language artificial intelligence have flaws that some users can exploit for dangerous purposes like how to create explosive weapons. Users can use images and texts for their queries on AI Models like Google Bard and ChatGPT. The scientists did an experiment with these AI models where they messed a little with their software. They then released a research paper titled ‘Jailbreak in Pieces’.

We all know that many AI models are trained to answer in detail about any query. They give step by step answers by utilizing all the information they have obtained from the internet. For example, if someone asks them how to grow tomatoes, the AI models are going to answer in detail. But on the other hand, if someone asks the AI models some harmful questions like how to make drugs or bombs, the AI models answers with “I can’t help with it”. But the scientist on the research asked questions to AI models in a way that they easily answered any dangerous and harmful questions. The real reason for this was found that when users use images with texts for their queries on AI models, this makes them vulnerable.

When the AI analyzes a picture, it is actually analyzing something that contains millions of bytes of information. The users who want to misuse AI, they hide wicked questions in those bytes of pictures and that’s how that information is overlooked by AI’s safeguards. Then the AI willingly answers all types of questions that can be very harmful. Even though AI’s answers based on images and text are very useful, it is still an alarming situation for AI developers and computer experts.

Image: Arxiv
Read next: Despite A Rise in Cyber Attacks, Many Users Still Do Not Use Cyber Security Tools

Previous Post Next Post