Google's SynthID watermarking system might have a problem. A developer named Aloshdenny says they've cracked it, and they've posted the code on GitHub for anyone to see.
The claim is pretty bold. According to Aloshdenny, all it took was 200 Gemini-generated images, some signal processing knowledge, and a lot of free time. No machine learning models, no insider access. Just pattern recognition and persistence.
If true, this matters because SynthID was supposed to be Google's answer to AI content provenance. The idea was to embed invisible watermarks in AI-generated images that would survive edits and compression. A way to track what's AI-made and what's not.
Aloshdenny's method allegedly works both ways. Strip watermarks from AI images, or add fake ones to human-created work. That second part is especially concerning for anyone relying on watermarks to verify content authenticity.
Google says the claim isn't accurate, but hasn't provided specifics about what Aloshdenny got wrong. The code is out there now, so security researchers will likely test it themselves.
This highlights a fundamental tension in AI watermarking. Any system that adds detectable patterns can theoretically be reverse-engineered. The question is whether the barrier is high enough to matter in practice.
For anyone building tools that rely on content verification, this is a reminder that watermarking alone probably isn't enough. You need multiple layers of authentication, especially as the stakes get higher.
