This attack was actually pointed out in the watermarking paper linked above. The researchers added an instruction to the prompt that switches letters like a Caesar Cipher. It lowers the quality of the output from the LLM but alters the "red list" enough for a watermark detection tool to fail at detecting the watermark.