AI Safety Bypass: JaiLIP Tricks AI with Invisible Pixel Changes

Microscopic changes in pixels, invisible to the human eye, can bypass the safety features of some artificial intelligence systems. Researchers at Florida International University found that an altered image, like a picture of a panda bear, can trick AI into creating harmful or policy-violating content. They developed an algorithm called JaiLIP, which determines the optimal pixel manipulation to bypass AI safeguards. When JaiLIP was used on a multimodal AI model called BLIP-2, it significantly increased the chance of the system generating unsafe responses. For example, an altered image of a stoplight caused the AI to provide instructions on how to disregard traffic signals without penalty. Using JaiLIP images nearly doubled the number of harmful responses from the tested AI models. This research shows that AI models interpret images differently than humans, seeing them as patterns of numbers and pixels. This highlights a critical vulnerability for businesses using AI.