AI Image Hack Hides Prompts to Steal User Data

Researchers have discovered a new method to extract sensitive information by embedding hidden prompts within images that are downscaled before being processed by AI systems. This AI Image Hack Hides commands that, once interpreted by a large language model (LLM), trigger unauthorized data access without detection. The attack manipulates the image processing pipeline, exploiting how AI models handle visual content before extracting embedded instructions.

The technique works by injecting prompts into image pixels, which remain effective even after the image is resized or compressed. These prompts can guide the LLM to retrieve or leak specific user data. The AI Image Hack Hides malicious commands in a way that standard security tools may overlook, raising concerns about the integrity of image-based AI workflows.

This evolving threat highlights the need for stronger safeguards in how AI platforms handle visual data. To learn more about how this attack works and its implications, read the full article below.
https://www.bleepingcomputer.com/news/security/new-ai-attack-hides-data-theft-prompts-in-downscaled-images/