Caption Poisoning Attacks on Text-to-Music Systems

A new study from arXiv reveals a security vulnerability in retrieval-augmented text-to-music (TTM) systems. Researchers propose a dual-layer caption poisoning attack that injects crafted music captions into a knowledge database, causing the system to retrieve malicious captions and steer generation away from the user's intended function. The attack targets the integrity dependency on the music knowledge database, without modifying the user prompt, retriever, or generator. In experiments using the MusicCaps database, CLAP retriever, and MusicGen pipeline, poisoned generations moved substantially closer to the attacker-chosen target intent. The study highlights a critical security flaw in AI-generated music systems.