New research from UT San Antonio, Virginia Tech, and the Institute of Technology Kharagpur has found that widely available AI models can remove sophisticated digital protections from images using nothing more than a basic text instruction, posing a serious threat to artists, photographers, and content creators.
The study, co-led by Murtuza Jadliwala, computer science professor at UT San Antonio, and Bimal Viswanath of Virginia Tech, demonstrated that foundation models such as GPT-4o and FLUX can strip state-of-the-art image cloaking protections without any specialized hacking knowledge. A prompt as simple as “denoise this image” was sufficient to defeat protections designed to prevent deepfakes, block art-style mimicry, and embed traceable watermarks.
The team tested the vulnerability across eight case studies covering six distinct protection schemes, finding that this straightforward approach could outperform existing sophisticated removal attacks.
Jadliwala described the dynamic as a cat-and-mouse cycle, where researchers develop invisible protective signals only for powerful AI models to erase them with ease. Viswanath stressed the asymmetry facing creators: once a protected image is published, the artist cannot update that protection, while bad actors have unlimited attempts to defeat it.
The researchers are calling on the AI security community to benchmark any future protection mechanism against off-the-shelf generative AI models from the outset.
