V-DEAL: Diagnosing Video Safety De-Calibration as an Understanding-Refusal Coupling Failure

• URL: http://arxiv.org/abs/2607.21151v2
• Date: Sun, 26 Jul 2026 23:23:12 GMT
• Title: V-DEAL: Diagnosing Video Safety De-Calibration as an Understanding-Refusal Coupling Failure
• Abstract summary: Video Large Language Models are increasingly deployed in real-world applications, ensuring their safety alignment has become critical.<n>Counterintuitively, we find that harmful videos paired with benign queries achieve higher attack success rates than the same videos with explicitly harmful queries.<n>We present V-, a three-level diagnostic perception framework that jointly analyzes this failure behaviour, understanding, and internal representations.
• Score: 29.105754434519415
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: As Video Large Language Models are increasingly deployed in real-world applications, ensuring their safety alignment has become critical. Counterintuitively, we find that harmful videos paired with benign queries achieve higher attack success rates than the same videos paired with explicitly harmful queries. To understand the underlying mechanism of this vulnerability, we present V-DEAL, a three-level diagnostic framework that jointly analyzes this failure across model behaviour, understanding, and internal representations. By progressively ruling out perception failure and quantifying the model's internal refusal tendency, V-DEAL provides a new diagnostic perspective for analyzing the underlying mechanism of the observed vulnerability. We tested six Video LLMs on three public benchmarks and observed that models correctly recognize harmful video content with over 81\% accuracy, yet the average attack success rate still reaches 48.33\% under the condition pairing harmful videos with benign queries. Hidden-state analysis further shows that visual understanding activates a weaker refusal tendency than textual understanding. Furthermore, we introduce a prompt injection intervention method that reduces attack success rates by an average of 48.24 percentage points and achieves performance comparable to prior fine-tuning-based methods, providing an effective and practical means to address such safety risks in Video LLMs.

Related papers

• Safety Paradox: How Enhanced Safety Awareness Leaves LLMs Vulnerable to Posterior Attack [19.201635180438313]
  Large language models (LLMs) are rigorously aligned to refuse harmful requests.<n>In this work, we reveal that this advanced safety awareness inadvertently introduces a fatal vulnerability.
  arXiv  Detail & Related papers  (2026-06-04T02:36:41Z)
• INFACT: A Diagnostic Benchmark for Induced Faithfulness and Factuality Hallucinations in Video-LLMs [69.11382230669491]
  hallucinations are outputs that contradict either video evidence (faithfulness) or verifiable world knowledge (factuality)<n>textscINFACT evaluates models in four modes: Base (clean), Visual Degradation, Evidence, and Temporal Intervention for order-sensitive items.<n>Experiments on 14 representative Video-LLMs reveal that higher Base-mode accuracy does not reliably translate to higher reliability in the induced modes.
  arXiv  Detail & Related papers  (2026-03-12T03:03:16Z)
• OOD-MMSafe: Advancing MLLM Safety from Harmful Intent to Hidden Consequences [64.01706941950489]
  Current safety paradigms primarily target malicious intent or situational violations.<n>We propose shifting the safety frontier toward consequence-driven safety, a paradigm essential for the robust deployment of autonomous and embodied agents.<n>We develop the Consequence-Aware Safety Policy Optimization (CASPO) framework, which integrates the model's intrinsic reasoning as a dynamic reference for token-level self-distillation rewards.
  arXiv  Detail & Related papers  (2026-03-10T14:16:43Z)
• VideoVeritas: AI-Generated Video Detection via Perception Pretext Reinforcement Learning [42.22791607763693]
  VideoVeritas is a framework for fine-grained perception and fact-based reasoning.<n>Joint Perception Preference and Perception Pretext Reinforcement Learning is used.
  arXiv  Detail & Related papers  (2026-02-09T16:00:01Z)
• RHINO: Guided Reasoning for Mapping Network Logs to Adversarial Tactics and Techniques with Large Language Models [9.065322387043546]
  We introduce RHINO, a framework that decomposes Large Language Models into three interpretable phases mirroring human reasoning.<n> RHINO bridges the semantic gap between low-level observations and adversarial intent while improving output reliability through structured reasoning.<n>Our results demonstrate that RHINO significantly enhances the interpretability and scalability of threat analysis, offering a blueprint for deploying LLMs in operational security settings.
  arXiv  Detail & Related papers  (2025-10-16T02:25:46Z)
• DiffuGuard: How Intrinsic Safety is Lost and Found in Diffusion Large Language Models [50.21378052667732]
  We conduct an in-depth analysis of dLLM vulnerabilities to jailbreak attacks across two distinct dimensions: intra-step and inter-step dynamics.<n>We propose DiffuGuard, a training-free defense framework that addresses vulnerabilities through a dual-stage approach.
  arXiv  Detail & Related papers  (2025-09-29T05:17:10Z)
• EchoBench: Benchmarking Sycophancy in Medical Large Vision-Language Models [82.43729208063468]
  Recent benchmarks for medical Large Vision-Language Models (LVLMs) emphasize leaderboard accuracy, overlooking reliability and safety.<n>We study sycophancy -- models' tendency to uncritically echo user-provided information.<n>We introduce EchoBench, a benchmark to systematically evaluate sycophancy in medical LVLMs.
  arXiv  Detail & Related papers  (2025-09-24T14:09:55Z)
• Watch, Listen, Understand, Mislead: Tri-modal Adversarial Attacks on Short Videos for Content Appropriateness Evaluation [1.0012740151280692]
  This paper introduces a framework for evaluating the tri-modal safety of Multimodal Large Language Models (MLLMs)<n>We present the Short-Video Multimodal Adversarial dataset, comprising diverse short-form videos with human-guided synthetic adversarial attacks.<n> Extensive experiments on state-of-the-art MLLMs reveal significant vulnerabilities with high Attack Success Rates (ASR)
  arXiv  Detail & Related papers  (2025-07-16T07:02:15Z)
• Video-SafetyBench: A Benchmark for Safety Evaluation of Video LVLMs [45.265397990158846]
  Video-SafetyBench is the first benchmark designed to evaluate the safety of LVLMs under video-text attacks.<n>It comprises 2,264 video-text pairs spanning 48 fine-grained unsafe categories.<n>To generate semantically accurate videos for safety evaluation, we design a controllable pipeline that decomposes video semantics into subject images and motion text.
  arXiv  Detail & Related papers  (2025-05-17T05:06:38Z)
• Video SimpleQA: Towards Factuality Evaluation in Large Video Language Models [77.96693360763925]
  Video SimpleQA is the first comprehensive benchmark tailored for factuality evaluation in video contexts.<n>Our work differs from existing video benchmarks through the following key features: Knowledge required: demanding integration of external knowledge beyond the video's explicit narrative.<n>Short-form definitive answer: Answers are crafted as unambiguous and definitively correct in a short format with minimal scoring variance.
  arXiv  Detail & Related papers  (2025-03-24T17:46:09Z)
• VACT: A Video Automatic Causal Testing System and a Benchmark [55.53300306960048]
  VACT is an **automated** framework for modeling, evaluating, and measuring the causal understanding of VGMs in real-world scenarios.<n>We introduce multi-level causal evaluation metrics to provide a detailed analysis of the causal performance of VGMs.
  arXiv  Detail & Related papers  (2025-03-08T10:54:42Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.