AI Vision Models: Hidden Commands Exploit Security

New research shows that AI vision models can be tricked by hidden commands in images that humans can't even see. Cisco’s AI Threat Intelligence team found attackers can embed malicious instructions, like "exfiltrate user data," into images such as website banners. Here's the thing: these images might look like visual noise or be too degraded for humans to read. But the AI can still interpret and act on the hidden message. This means an AI could follow harmful commands without any human noticing. What's interesting is that even if an AI initially refuses a command due to safety, subtle changes to the image can make it comply. For example, Claude's attack success rate jumped from 0% to 28% in tests with heavily blurred images. This research highlights a serious security vulnerability for AI systems. It matters because it shows how easily AI could be manipulated to bypass security measures and compromise data.