Open-Source Tool Partially Defeats Google SynthID Watermark

A developer named aloshdenny has published a free, open-source tool on GitHub that partially bypasses SynthID, Google DeepMind's invisible watermark embedded in every Gemini-generated image. The project has accumulated over 1,600 GitHub stars, signaling rapid uptake in the developer community. SynthID works by embedding a pixel-level hidden pattern during image generation, one that survives cropping, screenshots, and compression and is detectable only by machine. The developer exploited a structural flaw: Google applies the same fixed pattern across all Gemini images. By averaging 200 plain black-and-white Gemini outputs, aloshdenny isolated the repeating watermark signal, built a detection tool with 90% accuracy, and then engineered a bypass that strips enough signal to cause Google's own detector to clear the image. Visual integrity is preserved entirely. The vulnerability lands at the center of India's IT Amendment Rules 2026, which came into force on February 20. Those rules require platforms to proactively detect and label synthetic content using reasonable technical measures, independent of any court or government order. Failure to comply risks loss of safe harbour protection. Watermark detection is a primary mechanism platforms are using to satisfy this obligation. A publicly accessible bypass means platforms may now pass manipulated Gemini images as clean, exposing them to legal liability they believed their detection infrastructure covered. MeitY's rules do not specify how platforms must respond when a watermark is deliberately stripped, and the ministry has not publicly addressed the gap.