Multi-bit watermarking is used to embed a message, typically an account identifier, into large language model output so that harmful text can be traced back to its origin. A new arXiv preprint (2610.18359) describes an attack technique that can alter that embedded message without access to the watermarking key.
The attack allows the watermark to be changed to a different identifier, effectively pinning the output on someone else. Because the attacker does not need the key, the method poses a direct challenge to the reliability of watermark-based attribution.
The authors position this as a certified targeted displacement: the attacker can aim the altered watermark at a chosen target. The result is that watermarks, which are meant to provide accountability, can be turned into a tool for framing innocent users. The paper is a preprint and has not yet been peer-reviewed.