Blind PRNG Hijacking: Undetectable Attack Against LLM Watermarking Schemes

arXiv·medium signal

Shows that cryptographic watermarking schemes (KGW, Unigram, DipMark) can be defeated by hijacking the pseudo-random number generator they depend on, without changing the model output quality. The attack preserves text integrity — making it undetectable by current verification methods. Challenges a leading defense for AI content attribution.