A new attack method can erase invisible watermarks from AI-generated images while keeping the pictures intact.
Researchers built a system called LiBRA, short for Latent In-band Bidirectional Removal Attack, that targets the invisible watermarks many AI image generators embed to prove an image was machine-made. Earlier removal attacks tried to flip a watermark's decoded bits to the opposite pattern, but that inverted signal often stayed detectable, and pushing harder just degraded the image for no benefit. LiBRA instead nudges an image's representation in a public autoencoder's latent space until the watermark decoder's confidence drops to a coin-flip, pulling from whichever direction requires the smallest, least damaging change. The team also verifies success with a two-sided statistical test rather than assuming a confidence score alone proves the watermark is gone.
Watermarking is the main technical argument regulators and platforms point to when they promise users can tell real images from AI-generated ones. A method that quietly defeats that signal while preserving image quality undercuts that argument more than earlier attacks, which left visible damage or a detectable fingerprint behind.
Watermarking has always been a lock, not a vault - this is one more reminder that anyone who can see the key can eventually open the door.