Anthropic’s invisible code watermarking is already failing. Developers are bypassing these statistical fingerprints by using simple syntactic refactoring and encoding tricks.

By scrambling token probability biases, coders easily strip the AI’s identity markers. Because these watermarks rely on fragile, detectable patterns rather than cryptographic signatures, minor text transformations render them useless.

The cat-and-mouse game between AI security and human ingenuity has officially begun.