Challenge: Can you remove this watermark? I built a CLI watermarking tool with anti-AI defenses — try to break it.

Developer Vitruves has launched Firemark, an open-source command-line tool written in Rust designed to create watermarks that are exceptionally resistant to removal, particularly by AI-powered tools. The tool employs a multi-layered security stack inspired by physical document protection. This includes cryptographic filigrane patterns like guilloche and moiré effects, similar to those found on banknotes, and non-deterministic perturbation that ensures every rendered watermark is pixel-unique. This uniqueness prevents AI models from learning a consistent pattern to subtract. For PDFs, Firemark implements 'copy-paste poisoning,' embedding invisible scrambled text that corrupts extracted content.

A key defensive feature is adversarial prompt injection, where the watermark contains embedded text strips designed to confuse AI image inpainting and editing tools, potentially causing them to amplify the watermark instead of removing it. With 17 different watermark styles—from dense tiling to scattered mosaics—clean cropping becomes impractical. Vitruves has issued a public challenge, providing a sample watermarked document and inviting the community to attempt removal while preserving the original file's authenticity, layout, and metadata. The goal is to gather data on attack methodologies to further harden the tool against the growing threat of AI-based forgery and content stripping.