FIU Study: Altered Images Evade AI Safeguards - Mirage News

Researchers at Florida International University (FIU) have dropped a major bombshell: they've found that even tiny, almost invisible changes to images can trick AI systems into giving out harmful or incorrect information. Led by Associate Professor Hadi Amini and graduate assistant Md Jueal Mia, the team developed a method called JaiLIP, which nearly doubled the number of unsafe responses from AI models like BLIP-2 in tests. This isn't just a technical glitch; it's a serious wake-up call for businesses, especially smaller ones, that are increasingly relying on AI-powered tools for customer service and other critical tasks. These 'image-based hacks' are a type of adversarial attack that exploits how AI 'sees' patterns in pixels, not meaning, potentially opening doors for cyberattacks and eroding user trust. Experts in adversarial AI have warned for a while that such subtle manipulations, called perturbations, can lead to AI systems making dangerous misclassifications. Moving forward, the FIU team is focused on staying ahead of the bad actors by finding more vulnerabilities and building stronger defense mechanisms. For businesses, the takeaway is clear: don't just set up AI and forget about it. They need to limit sensitive information fed into AI, restrict access, and continuously evaluate the security measures and guardrails of their AI tools through ongoing adversarial testing to keep systems safe and trustworthy.