Search papers, labs, and topics across Lattice.
This paper investigates the impact of benign fine-tuning on the safety alignment of Audio LLMs, revealing that it significantly degrades safety, increasing Jailbreak Success Rates (JSR) to as high as 87.12%. By decomposing embedding-space proximity into semantic, acoustic, and mixed axes, the study demonstrates that the dominant vulnerability axis and the relative risk of audio versus text fine-tuning are architecture-dependent. The authors propose and validate two effective defenses: filtering training data based on distance from harmful embeddings and using a textual system prompt at inference, both of which reduce JSR to near-zero.
Benign fine-tuning can turn your safe Audio LLM into a jailbreaking machine, and the way it breaks depends on the model's architecture.
Prior work shows that fine-tuning aligned models on benign data degrades safety in text and vision modalities, and that proximity to harmful content in representation space predicts which samples cause the most damage. However, existing analyses operate within a single, undifferentiated embedding space -- leaving open whether distinct input properties drive the vulnerability differently. Audio introduces a structurally richer problem: a benign sample can neighbor harmful content not only through what is said but through how it sounds, even when its words are entirely innocuous. We present the first systematic study of benign fine-tuning safety in Audio LLMs, evaluating three state-of-the-art models with a proximity-based filtering framework that selects benign audio by embedding-space distance to harmful content. By decomposing proximity into semantic, acoustic, and mixed axes using external reference encoders alongside each model's own internal encoder, we show that benign fine-tuning elevates Jailbreak Success Rate (JSR) from single digits to as high as 87.12%. Crucially, the dominant vulnerability axis and the relative risk of audio versus text fine-tuning are both architecture-conditioned -- determined by how each model's encoder and projector transform audio into the LLM's input space. We propose two defenses: filtering training data to maximize distance from harmful embeddings, and a textual system prompt at inference, both reducing JSR to near-zero without architectural modification. Our mechanistic analysis on two architectures reveals that fine-tuning selectively suppresses the late-layer refusal circuit while the frozen encoder preserves representations, and that even the suppression pattern is architecture-conditioned, mirroring the behavioral asymmetries across modalities. Safety degradation from benign fine-tuning is a qualitatively distinct risk in Audio LLMs.
