EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Hijacking the “Safety Bifurcation Point”: Coupled Path Hijacking Attacks on Multimodal Vision–Language Models

Ranyi Peng, Jingfu Bao

Electronics · 16.07.2026 · 15 (14) · 3143

DOI: 10.3390/electronics15143143 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Electronics, online erschienen am 16.07.2026. Verfasst von Ranyi Peng und 1 weiteren Autor:innen. Die Arbeit nennt 47 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

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.

Zitieren

APA

Peng, R., & Bao, J. (2026). Hijacking the “Safety Bifurcation Point”: Coupled Path Hijacking Attacks on Multimodal Vision–Language Models. Electronics, 15(14), 3143. https://doi.org/10.3390/electronics15143143

Vancouver

Peng R, Bao J. Hijacking the “Safety Bifurcation Point”: Coupled Path Hijacking Attacks on Multimodal Vision–Language Models. Electronics. 2026;15(14):3143. doi:10.3390/electronics15143143

BibTeX
@article{Peng2026,
  author = {Peng, Ranyi and Bao, Jingfu},
  title = {Hijacking the “Safety Bifurcation Point”: Coupled Path Hijacking Attacks on Multimodal Vision–Language Models},
  journal = {Electronics},
  year = {2026},
  volume = {15},
  number = {14},
  pages = {3143},
  doi = {10.3390/electronics15143143}
}
Kostenlos starten →