The Hack of SynthID or How Not to Expose Your Generations

TG AI News·April 13, 2026 at 5:49 PM·
Trusted Source
The acclaimed SynthID from Google has fallen victim to simple reverse engineering. Let me remind you that this is just a clever layer of structured noise. The eye cannot see it, and compression does not kill it, but on a pure white or black background, this filter is instantly exposed. So, a skilled individual was able to extract the noise pattern by running it through Nano Banana Pro on completely black and white backgrounds, and then, through spectral analysis, capturing the frequency coordinates and packing them into a dictionary. The tool erases the watermark with surgical precision — breaking 90 percent of the protection and yielding 43 dB PSNR, meaning the visual quality does not suffer at all. The only nuance is that the carrier frequencies change with the resolution of the image, so the developer is now calling on the community to submit pull requests and enrich the dictionary database. It seems that any attempts by corporations to secretly mark their content are still being dismantled by an open-source Python script in no time.
The Hack of SynthID or How Not to Expose Your Generations | AI News | AIventa