Exploit that image, AI

Based on this article by Deeba Ahmed at Hackread.

Cybersecurity researchers have discovered a new vulnerability where hackers can hide malicious commands within seemingly ordinary images to exploit AI chatbots. This "image scaling attack" works by embedding instructions that only become visible to the AI after it automatically shrinks a large image for processing.

The researchers demonstrated the threat on platforms like Google Gemini, showing how a single image could command the AI to steal a user's private calendar data. To prevent such attacks, experts recommend that AI systems require explicit user confirmation before executing any sensitive tasks prompted from an image.

Check out the article.

Related Stories