AIDigital Trendsabout 8 hours ago
Hidden photo trick bypasses AI safety rules
1 min read
Florida International University researchers built JaiLIP, a method that alters pixels in photos invisibly to trick AI chatbots into ignoring safety rules. Testing on BLIP-2 nearly doubled harmful responses. A modified stoplight photo made the model explain how to run a red light without a ticket.
Level
Hype check
Tap to vote and see what everyone thinks.
