Answers are grounded in this paper and cite their sources.
This randomized trial reveals vulnerabilities in safety mechanisms of LVLMs, suggesting new defensive strategies for AI systems.
The generalization of safety mechanisms from text-only LLMs to Large Vision–Language Models (LVLMs) remains poorly understood, especially concerning their vulnerability to typographic adversarial prompts. In this paper we ask where, inside a model, the safe-versus-unsafe decision is causally made. Focusing on decoder-only LVLMs, we show that safety is encoded as a non-linear, distributed representation that bifurcates early at a distinct Safety Bifurcation Point (SBP)—e.g., Layer 9 in LLaVA-1.5 and Layer 5 in Qwen-VL—and we expose a Correlation–Causality Gap: late layers are highly linearly separable yet causally inert. To leverage this insight we introduce a multistage, causal (interpretability-guided) framework to locate and verify safety representations, and we propose CPH (Coupled Path Hijacking), a white-box attack that jointly patches MLP and self-attention activations at the SBP to deterministically hijack the safe computational path. On the two models for which we report full tabulated results, CPH attains high oracle-scored attack success rates of 99.43% (LLaVA-1.5) and 98.57% (Qwen-VL); we additionally observe 99.71% on InstructBLIP-Vicuna, whose Vicuna backend is likewise decoder-only. We contrast this decoder-only vulnerability with encoder–decoder designs (e.g., a T5 backend or a Q-Former front-end), in which cross-modal grounding precedes decoding; we advance the greater robustness of such architectures as an architectural hypothesis for future validation, not as a tabulated result of this study. These findings indicate that VLM safety is architecture-dependent and motivate architecture-aware defenses for trustworthy multimodal AI.
No takes yet. Share an insight, caveat, or question.
Peng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: