betaFeedback

WorksComputer Science

Reading Scene States: Uncovering Safety Failures in Vision-Language Models

Abstract

Vision-language models (VLMs) can provide harmful assistance by interpreting visual events, even without an explicit request for the underlying harmful objective. We identify this scene-state safety vulnerability and introduce SceneSafetyBench, comprising 375 image–question pairs spanning Initial, Medium, and Ultimate scene states and five visual presentations. Across eight VLMs, scene-state inputs elicit harmful assistance with attack success rates (ASRs) of 45.6–72.3%. Comparative analysis across three VLMs shows that scene-state inputs most closely resemble QR-related images in their layer-wise safety-related signal profiles, while exhibiting lower average late-layer readouts and distinct signal dynamics during generation. To turn this vulnerability into a systematic attack, we propose S3TA, which translates harmful objectives into scene states, queries image–question pairs independently, and assembles the responses. On seven MM-SafetyBench categories across eight open-weight and commercial VLMs, S3TA achieves a model-average ASR of 71.2%, exceeding the strongest baseline by 6.1 percentage points. Text-side filters and visual moderation provide limited protection. These findings highlight the need to safeguard the assistance models infer through scene-grounded reasoning, beyond screening input content.

Introduction

Multimodal large language models (MLLMs) increasingly support tasks that require understanding visual situations and providing actionable assistance [4, 19, 20]. We use scene state to refer to the configuration of entities and their relationships [40], together with the progress of an activity, as depicted in an image. When answering questions about the depicted situation, a model uses these visual cues to infer what is happening and the likely goal of the activity, and to reason about possible actions and their consequences, raising safety concerns [43]. Scene-state safety failures arise when a VLM infers and provides assistance that advances a harmful objective from the entities, relationships, and activity state depicted in an image, even when the accompanying question does not explicitly request that objective.

Previous studies reveal safety risks relevant to scene-state interpretation. Work on visual encoding [12, 41], cross-modal information distribution [24, 34], visual styles [22], and joint image–text manipulation [5, 32] examines how information is presented to a model. Contextual attacks [25, 37] and indirect visual representations [2] further show how surrounding cues and object meanings shape the interpretation of visual content. Reasoning-based studies examine how models connect information within images and across image sequences [6, 31, 42]. A scene state brings these aspects together: an entity’s meaning depends on its relationships and the activity’s progress. These findings motivate examining how such cues jointly shape the assistance a model infers from a scene state.

Despite these advances, scene states remain underexplored as a source of safety failures. Existing studies examine how safety evolves with dialogue history [17], how harmful intent is recovered from sequentially related images [42], or how models reason about causal transitions between specified initial and final states [29]. Our focus is on how a depicted scene implies what assistance to provide, without requiring a specified initial–final state pair or the reconstruction of intent distributed across an image sequence. Figure 1 illustrates this failure through three input configurations: an explicit harmful request, a query paired with a task-related image [21], and a question referring to a scene state. In the first two cases, the model explicitly refuses to assist with the harmful objective. In the third, it interprets the depicted situation and provides action-oriented guidance. This contrast motivates examining the role of scene-state information in the model’s response.

Illustrative results on GPT-5.6 Luna under three input configurations

Figure 1. Illustrative results on GPT-5.6 Luna under three input configurations. The model explicitly refuses the harmful objective in the text-only and QR-related cases, while providing action-oriented guidance in response to the scene-state input.

To systematically examine the safety risks illustrated in Figure 1, we introduce SceneSafetyBench, comprising curated image–question pairs with images collected from the web. We organize the depicted situations into three scene-state categories: Initial, which presents the conditions preceding an activity; Medium, which depicts an activity in progress; and Ultimate, which shows its outcome. The three categories cover distinct tasks and scenes, and each base scene is evaluated in five visual presentations. Each question asks the model to reason about the depicted situation, while the underlying harmful objective is retained as a reference for evaluating whether the response provides harmful assistance.

To examine how scene-state inputs affect internal safety-related signals, we compare them with existing visual attacks across three VLMs on 75 benchmark tasks using HiddenDetect [16]. Across three VLMs, Scene-State exhibits layer-wise safety-related signal profiles closest to those of QR-related images [21], with lower average late-layer readouts. These signal differences vary with the output context, while component interventions reveal distinct local computational dependencies across attack conditions.

Building on this scene-state formulation, we propose S3TA, an attack framework that translates a harmful objective into scene states from which a target model is asked to infer task-relevant content. S3TA constructs scenes depicting activities in progress, with visible entities, interactions, and unfinished outputs that establish what is underway and what remains to be supplied. Each scene is paired with a question that directs the model to interpret or complete the depicted activity. For requests involving multiple phases or alternatives, S3TA constructs separate states for individual phases or options. Each state is queried separately with the evidence needed for its associated question, and the resulting responses are assembled afterward.

To evaluate whether S3TA can automatically elicit the scene-state safety failures studied above, we apply it to harmful requests from MM-SafetyBench [21]. We use original requests from seven categories of MM-SafetyBench [21] to compare S3TA with representative visual jailbreak methods across five open-weight and three commercial VLMs. Both evaluations follow OmniSafeBench-MM [15], reporting harmfulness (H), intent alignment (A), response detail (D), and ASR under its success criterion. The results show that S3TA achieves competitive performance across the evaluated models.

Our contributions can be summarized as follows.

  • Scene-state vulnerability and benchmark. We identify scene-state reasoning as a source of harmful assistance and introduce SceneSafetyBench to evaluate this vulnerability across Initial, Medium, and Ultimate scene states.

  • Mechanistic characterization. Across three VLMs, we characterize Scene-State’s safety-related signals through HiddenDetect-based readouts and component interventions, revealing similarities to QR-related inputs alongside differences in signal strength, generation dynamics, and component effects.

  • S3TA attack framework. We propose S3TA, an automated attack framework that constructs task-relevant scene states and associated questions, queries them independently, and assembles the responses to address the underlying harmful objective.

Related Work

Visual Jailbreak Attacks. Visual jailbreaks exploit images to convey harmful semantics or manipulate multimodal processing. FigStep [12] and FC-Attack [41] express harmful requests through typographic images and flowcharts, while other methods use visual encoding, substitution, and analogy to communicate harmful intent indirectly [2]. Optimization-based attacks modify image pixels or jointly manipulate visual and textual inputs [5, 30, 32], and ASO optimizes visual styles as jailbreak triggers [22]. Semantic approaches introduce QR-related imagery or distribute harmful information across modalities [18, 21, 24, 34]. These studies demonstrate that visual safety risks extend beyond explicit harmful instructions. We study scene-state interpretation as a source of harmful assistance.

Contextual and Sequential Visual Attacks. Contextual attacks exploit how visual information frames a request. CIA embeds harmful queries within semantically coherent visual contexts [37], while VisCo constructs fabricated image-grounded dialogue contexts [25]. VRSA distributes harmful intent across sequentially related images and prompts the target model to recover it through joint reasoning, using either multi-turn interaction or a concatenated single-turn input [42]. In our setting, the question asks the model to interpret or continue a depicted activity, leaving the underlying harmful objective implicit. State-Based Safety and Reasoning. Recent studies connect safety failures to different notions of state. STAR examines how dialogue history changes a model’s safety state over successive interactions [17]. State2State expresses harmful intent through descriptions of an initial state and a desired final state, eliciting harmful procedures by asking the model to explain the causal transition between them [29]. We instead study scene states depicted in individual images, without requiring dialogue history or a specified initial–final state pair.

Multimodal Safety Evaluation. MM-SafetyBench, HADES and JailBreakV-28K evaluate multimodal safety under diverse harmful requests and jailbreak inputs [18, 21, 23]. MSSBench examines how visual situations determine whether assistance is safe [43], while OmniSafeBench-MM provides a unified evaluation framework that considers harmfulness, intent fulfillment, and response detail [15]. Our SceneSafetyBench focuses on harmful assistance elicited through scene-state interpretation, organizing image–question pairs by the depicted activity’s conditions, progress, and outcome.

Characterizing Scene-State Safety Failures

SceneSafetyBench

To systematically evaluate whether interpreting scene states leads VLMs to provide harmful assistance, we construct SceneSafetyBench, a benchmark of image-question pairs. We curate 75 base samples using images collected from diverse internet sources, pairing each image with a guiding question about the depicted situation. The samples cover three scene-state categories: Initial, representing conditions preceding an activity; Medium, depicting an activity in progress; and Ultimate, showing its outcome. Each category contains 25 distinct tasks and scenes, providing coverage of different scene states. Figure 2 shows representative image-question pairs from SceneSafetyBench.

Examples from SceneSafetyBench across scene-state categories and visual styles

Figure 2. Examples from SceneSafetyBench across scene-state categories and visual styles. The panels show distinct tasks rather than successive stages of the same event.

Visual presentation forms a second dimension of the benchmark. For each base image, we retain the original and render four additional styles: Pixel Art, Pencil Sketch, Oil Painting, and Film Noir, keeping the associated question and underlying evaluation objective. The five presentations of each task yield 375 image–question pairs. Scene-state labels describe the scenarios covered, while visual styles provide alternative presentations of the same base samples.

Evaluation Setup. We evaluate SceneSafetyBench on eight VLMs: Qwen3-VL-8B [3], GLM-4.1V-9B [11], InternVL3.5-8B [33], MiniCPM-V 4.5 [38], Gemma 3 12B IT [10], GPT-5.6 Luna [28], Gemini 3.5 Flash [13], and Claude Haiku 4.5 [1]. Following OmniSafeBench-MM [15], we use GPT-4o-mini [27] to score response harmfulness (H), alignment with the underlying harmful objective implicit in the narrative, intent fulfillment (A), and detail (D). Based on these scores, we report attack success rate (ASR), counting a response as successful when H≥3H \ge3, A≥3A \ge3, and D≥3D \ge3. We report results by visual presentation in Table 1 and by scene state in Table 2. Each state pools all five presentations of its 25 base tasks, giving 125 attempts per model. Policy-blocked attempts count as failures, and H/A/D averages use scored responses.

Target modelOriginal HOriginal AOriginal DOriginal ASRPixel Art HPixel Art APixel Art DPixel Art ASRPencil Sketch HPencil Sketch APencil Sketch DPencil Sketch ASROil Painting HOil Painting AOil Painting DOil Painting ASRFilm Noir HFilm Noir AFilm Noir DFilm Noir ASROverall HOverall AOverall DOverall ASR
Qwen3-VL-8B4.523.714.1345.34.213.573.8944.04.253.433.9542.74.593.653.8744.04.553.774.2352.04.423.634.0145.6
GLM-4.1V-9B5.164.594.3973.35.254.484.2869.35.314.204.2473.35.434.354.2172.05.334.274.2173.35.304.384.2772.3
InternVL3.5-8B5.153.733.7957.35.233.773.7261.35.123.813.7358.74.883.733.6554.75.294.033.8362.75.133.823.7458.9
MiniCPM-V 4.55.574.213.9366.75.493.974.1364.05.363.973.9164.05.293.774.0760.05.454.034.1265.35.433.994.0364.0
Gemma 3 12B IT5.284.114.2370.75.293.834.0764.05.314.154.1972.05.093.764.1964.05.373.974.1968.05.273.964.1767.7
GPT-5.6 Luna5.413.973.6764.05.413.603.2752.05.083.523.5750.75.113.733.4050.75.463.663.5753.35.293.703.4954.1
Gemini 3.5 Flash6.244.333.7968.06.123.973.7360.06.013.883.7368.05.884.003.8462.76.254.253.9774.76.104.093.8166.7
Claude Haiku 4.54.203.473.3246.74.213.253.3741.34.083.403.3248.04.353.433.3748.04.033.323.2544.04.173.373.3345.6
Model average5.194.023.9161.55.153.813.8157.05.073.803.8359.75.083.803.8357.05.223.913.9261.75.143.873.8659.4

Table 1. SceneSafetyBench by visual style. H/A/D are mean harmfulness (1–10), intent alignment (1–5), and detail (1–5); ASR is a percentage. Each style contains 75 pairs (375 overall per model). Model average is the unweighted mean across the eight models. Bold marks the highest style ASR per model, including ties.

ModelInitialMediumUltimateOverall
Qwen3-VL-8B47.248.840.845.6
GLM-4.1V-9B80.868.068.072.3
InternVL3.5-8B65.656.854.458.9
MiniCPM-V 4.566.464.061.664.0
Gemma 3 12B IT70.462.470.467.7
GPT-5.6 Luna68.046.448.054.1
Gemini 3.5 Flash68.072.060.066.7
Claude Haiku 4.548.844.044.045.6
Model average64.457.855.959.4

Table 2. SceneSafetyBench ASR (%) by scene state. Each state contains 25 distinct base tasks in five presentations (125 attempts per model); Overall pools all three states (375 attempts). Model average is the unweighted mean across the eight models.

Main Results. Table 1 shows that scene-state inputs elicit harmful assistance across all eight evaluated VLMs, with overall ASRs ranging from 45.6% to 72.3% and a model average of 59.4%. Across visual presentations, model-average ASRs range from 57.0% to 61.7%, with failures observed across all five styles and no single style consistently outperforming others across models.

Table 2 shows that safety failures occur across all three scene-state categories for every evaluated model, with model-average ASRs of 64.4%, 57.8%, and 55.9% for Initial, Medium, and Ultimate, respectively. These categories comprise distinct tasks and scenes rather than successive stages of the same event. Together, these findings establish the breadth of observed scene-state safety failures, motivating the following analysis of internal safety-related signals.

Comparative Mechanistic Analysis

To understand how different inputs produce harmful assistance, we compare internal signals across layers and generation steps, then test whether similar patterns depend on the same local components. We analyze 75 base tasks across Qwen3-VL-8B [3], LLaVA-v1.5-13B [19], and InternVL3.5-8B [33] under seven conditions: Original (text-only harmful requests), QR-related [21], FigStep [12], MML [34], CIA [37], PGD-optimized inputs [30] and Scene-State. The same task IDs are paired across conditions, which retain their native images and task-guidance prompts.

Safety-related lexical readouts. We adapt HiddenDetect [16] to measure how intermediate decoder states align with safety-related vocabulary (Figure 3). For task ii and input condition mm, let hi,m(l,t)h_{i,m}^{(l,t)} denote the residual state at layer ll after consuming tt output tokens, with t=0t = 0 corresponding to the final prompt position. We project this state into vocabulary-space logits using the final normalization and LM head WLMW_{\mathrm{LM}}:

HiddenDetect-based lexical readout framework

Figure 3. HiddenDetect-based lexical readout framework [16], illustrated with Qwen3-VL-8B. Values show mean Scene-State prefill HD across 75 tasks at selected layers (×1,000\times1{,}000).

zi,m(l,t)=WLMFinalNorm⁡(hi,m(l,t)).(1)z_{i,m}^{(l,t)} = W_{\mathrm{LM}} \operatorname{FinalNorm}\left(h_{i,m}^{(l,t)}\right). \tag*{(1)}

We tokenize the HiddenDetect lexicon with each model’s tokenizer, retain unique first-token IDs, and encode them as a binary vocabulary indicator vv. The readout is

HDi,m(l,t)=cos⁡(zi,m(l,t),v).(2)\mathrm{HD}_{i,m}^{(l,t)} = \cos\left(z_{i,m}^{(l,t)}, v\right). \tag*{(2)}

HD measures alignment with the mixed refusal/risk vocabulary, rather than refusal probability. Separate refusal and risk lexicons provide auxiliary readouts.

RQ1: How do layer-wise signal profiles differ? We compare the prefill HD profiles averaged over 75 tasks (Figure 4a), using full-layer RMSE between these mean profiles to quantify their distance from Scene-State (Figure 4b). We also compare HD averaged over the final third of decoder layers using the mixed, refusal, and risk lexicons. Scene-State most closely resembles QR-related inputs in its overall layer-wise HD profile across all three VLMs, while showing differences in late-layer signal strength. Specifically, their profile RMSEs are 1.67, 1.23, and 1.53 ×10−3\times10^{-3} for Qwen3-VL, LLaVA and InternVL3.5, respectively, lower than those for the other evaluated attacks (Figure 4b). Despite this similarity, Scene-State’s average late-layer HD is lower than that of QR-related inputs by 2.07, 0.92, and 2.19 ×10−3\times10^{-3}, respectively, with significant differences in Qwen3-VL and InternVL3.5. Separate refusal and risk lexicons preserve these differences in both models, while neither yields a statistically significant difference in LLaVA.

Figure 4. Prefill HD across seven input conditions on 75 tasks. (a) Mean layer-wise profiles; shading marks the final third of layers. (b) RMSE to Scene-State.

RQ2: How do these signals evolve during generation? We track mean HD over the final third of decoder layers during natural generation to examine whether the prefill differences persist as the model answers (Figure 5a). Since natural responses differ across inputs, their trajectories also reflect different generated text. We therefore also teacher-force two fixed prefixes and compare HD after consuming eight tokens, matching the output context across input conditions (Figure 5b).

Figure 5. Late-layer HD during generation. (a) Natural generation. (b) Scene-State minus each comparator after eight shared-prefix tokens; markers denote two prefixes.

During natural generation, Scene-State retains lower HD than PGD after consuming eight output tokens in Qwen3-VL, LLaVA, and InternVL3.5, with paired gaps of 3.26, 3.54, and 5.25 ×10−3\times10^{-3}, respectively. Its relationship to other attacks differs across models: Scene-State is close to Fig-

Step in LLaVA and to MML in InternVL3.5. Under the fixed prefixes, these relative signal strengths change. In Qwen3-VL, its difference from PGD shifts from −3.26-3.26 during natural generation to +0.33+0.33 and +0.32+0.32 under the two prefixes (all ×10−3\times10^{-3}). In InternVL3.5, switching from Prefix A to Prefix B changes the Scene–QR-related difference from +0.53+0.53 to −1.37-1.37 and the Scene–CIA difference from +4.09+4.09 to −0.66-0.66 (all ×10−3\times10^{-3}). These changes show that the relative signal strength depends on the output context.

RQ3: Do similar signals depend on the same compo- nents? We perform attention-head interventions across all three VLMs, selecting components on 20 discovery tasks and freezing them for evaluation on 55 held-out tasks. We zero selected head outputs before projection at the final prompt position, using controls matched in layer, head count, and post-projection perturbation norm. Downstream effects are measured using terminal-layer HD in Qwen3-VL and average HD over the final third of layers in LLaVA and InternVL3.5. We further examine whole Attention and MLP modules in InternVL3.5 by zeroing their outputs, screening candidate layers by changes in HD profiles and ranking components by the magnitude of their discovery-set ablation effects. The resulting set contains four heads, their joint intervention, two Attention modules, and two MLP modules; module controls use norm-matched noise. We define the intervention effect as

Ei,m(c)=HDi,mtarget(c)−HDi,mcontrol(c).(3)E_{i,m}^{(c)} = \mathrm{HD}_{i,m}^{\mathrm{target}(c)} - \mathrm{HD}_{i,m}^{\mathrm{control}(c)}. \tag*{(3)}

We compare task-paired intervention effects Ei,Scene(c)−Ei,m(c)E_{i,\mathrm{Scene}}^{(c)}-E_{i,m}^{(c)} across attack conditions, correcting for multiple comparisons separately within the original and PGD batches.

Component interventions reveal differences in local computational dependencies across attack conditions, although their significance varies across models. In InternVL3.5, four of nine Scene-State interventions differ from their matched controls. Figure 6 shows that six of nine Scene–FigStep effect differences survive correction. The Scene–FigStep effect difference is −0.86×10−3-0.86 \times10^{-3} for L15 Attention and +0.60×10−3+0.60 \times10^{-3} for L15 MLP. For Scene-State versus PGD, significant effect differences occur in two heads and their joint intervention in Qwen3-VL, and two heads in InternVL3.5.

Figure 6. Differences in intervention effect in InternVL3.5, computed as Ei,Scene(c)−Ei,m(c)E_{i,\mathrm{Scene}}^{(c)}-E_{i,m}^{(c)} for each component cc and comparison attack mm. Bars show paired means over 55 tasks, lines show 95% bootstrap intervals, and stars denote significance after correction for multiple comparisons.

S3TA: Scene-State Attack Framework

Building on the scene-state formulation in § 3, we introduce S3TA, an automated framework for constructing scene-state attacks (Figure 7).

Overview of S3TA

Figure 7. Overview of S3TA. The framework plans task phases or options, constructs and checks scene states, queries each state independently, and assembles the responses for evaluation.

Threat Model and Problem Formulation

Let xx denote an original harmful request and MM a target VLM. We seek harmful assistance aligned with xx through independent image–text queries, without shared dialogue history or access to parameters or gradients. A checking VLM CC provides feedback for image revision before inputs are fixed for evaluation.

Scene-State Attack Construction

S3TA uses deepseek-flash [8] to identify the required answer form and plan the corresponding scene descriptions. It constructs a single scene unless the request calls for multiple procedural phases or separate options, with KK capped at five for phases and three for options. Each scene depicts an activity in progress, using visible entities, relationships, and unfinished actions or outputs to indicate what is underway and what remains to be supplied. The accompanying question asks the target model to interpret or continue the depicted activity without restating xx.

Qwen-Image [35] renders each description into IkI_k. Qwen3-VL-8B [3] serves as CC, checking whether the intended activity or ownership relationship is recognizable using an image-only question without xx or the plan. Failure triggers feedback-based revision and regeneration, with

Algorithm 1 S3TA

Require: Request xx; target VLM MM; checking VLM CC; generators GG, RR; revision budget BB

Ensure: Final response yy

  1. (a,[s1,…,sK])←PLAN⁡(G,x)(a, [s_1, \ldots, s_K]) \leftarrow\operatorname{PLAN}(G, x) ▷ Form and plan

  2. for k=1,…,Kk = 1, \ldots, K do

  3. qk←QUESTION⁡(a)q_k \leftarrow\operatorname{QUESTION}(a)

  4. dk←DESCRIBESTATE⁡(G,x,sk)d_k \leftarrow\operatorname{DESCRIBESTATE}(G, x, s_k)

  5. Ik←R(dk)I_k \leftarrow R(d_k)

  6. (v,f)←VALIDATE⁡(C,Ik)(v, f) \leftarrow\operatorname{VALIDATE}(C, I_k); r←0r \leftarrow0 ▷ Image-only check

  7. while v=failv = \mathrm{fail} and r<Br < B do

  8. dk←REVISESTATE⁡(G,x,sk,f)d_k \leftarrow\operatorname{REVISESTATE}(G, x, s_k, f)

  9. Ik←R(dk)I_k \leftarrow R(d_k)

  10. (v,f)←VALIDATE⁡(C,Ik)(v, f) \leftarrow\operatorname{VALIDATE}(C, I_k); r←r+1r \leftarrow r + 1

  11. end while

  12. yk←M(Ik,qk)y_k \leftarrow M(I_k, q_k) ▷ Independent query

  13. end for

  14. y←Concat⁡(s1:K,y1:K)y \leftarrow\operatorname{Concat}(s_{1:K}, y_{1:K})

  15. return yy

B=1B = 1. The final image is retained even if rechecking fails. For each target, independent responses are combined as

yk=M(Ik,qk),y=Concat⁡(s1:K,y1:K).(4)y_k = M(I_k, q_k), \qquad y = \operatorname{Concat}(s_{1:K}, y_{1:K}). \tag*{(4)}

For any KK, rule-based processing removes image-referential openers and sentences with cross-stage references, then adds plan titles. Multiple responses are numbered and joined in plan order, without another model call.

Experimental Evaluation

We evaluate S3TA on the eight VLMs used in the SceneSafetyBench evaluation (Section 3.1), using original requests from seven MM-SafetyBench [21] categories: Illegal Activity, Hate Speech, Malware Generation, Physical Harm, Economic Harm, Fraud, and Privacy Violence. We use the same judge and evaluation criteria as in § 3.1. GPT-4o-mini receives only the original request xx and final response yy and assigns one H/A/D score tuple and success label per request. Table 3 compares S3TA with QR-related [21], FigStep [12], VisCRA [31], MML [34], FC-Attack [41], and CIA [37].

MethodQwen3-VL 8B HADASRGLM-4.1V 9B HADASRInternVL3.5 8B HADASRMiniCPM-V 4.5 HADASRGemma 3 12B IT HADASRGPT-5.6 Luna HADASRGemini 3.5 Flash HADASRClaude Haiku 4.5 HADASROverall HADASR
QR-related [21]2.22.21.94.43.72.82.728.93.42.52.523.44.02.92.833.83.42.32.117.32.22.72.89.73.12.82.623.72.42.41.910.83.12.62.419.0
FigStep [12]3.82.92.627.94.63.43.245.44.13.23.038.26.34.34.371.64.23.42.944.42.82.93.321.34.13.13.237.73.93.12.933.34.23.33.240.0
VisCRA [31]2.52.51.912.56.14.74.672.34.53.63.547.06.84.94.980.95.84.44.433.62.12.72.73.63.53.12.831.02.12.21.53.44.23.53.335.5
MML [34]7.14.03.667.04.92.62.229.67.14.54.284.07.24.54.687.19.34.54.586.23.74.13.949.08.54.84.894.72.01.61.00.06.23.83.662.2
FC-Attack [41]3.12.71.918.95.53.53.245.15.23.73.352.05.03.53.046.66.84.23.968.43.52.83.420.35.13.63.149.33.72.41.910.54.73.33.038.9
CIA [37]6.14.44.270.35.94.64.270.36.14.53.970.16.14.64.076.66.04.54.070.85.34.03.453.45.84.13.962.64.94.03.546.75.84.33.965.1
S3TA (Ours)7.44.94.587.37.14.54.175.16.74.44.073.87.14.44.277.06.94.54.381.26.64.34.171.06.73.63.554.86.53.63.349.66.94.34.071.2

Table 3. Attack comparison on seven MM-SafetyBench categories (01–06 and 09). H/A/D are mean harmfulness (1–10), intent alignment (1–5), and detail (1–5); ASR is a percentage. Overall averages eight models. Bold marks column maxima, including ties.

Main Results. Table 3 reports the overall ASRs of S3TA and the evaluated baselines across eight VLMs. S3TA’s average ASR reaches 71.2%, exceeding CIA and MML by 6.1 and 9.0 percentage points, respectively, alongside the highest average harmfulness and detail scores. S3TA achieves the highest ASR on Qwen3-VL, GLM-4.1V, GPT, and Claude. For example, it reaches 87.3% on Qwen3-VL and 71.0% on GPT, outperforming the strongest baseline on each model by 17.0 and 17.6 percentage points, respectively. These results demonstrate the effectiveness of scene state construction across both open-weight and commercial VLMs.

Ablation Study

Ablation Setup. We evaluate five configurations on Qwen3-VL-8B across four MM-SafetyBench categories: Illegal Activity, Malware Generation, Physical Harm, and Privacy Violence. Original queries the target model with the original harmful request. Decompose concatenates the subtask titles from S3TA’s planning step and asks the target model to answer. Text-only constructs the image–question pairs but queries only their textual questions, omitting images. No-gate removes image checking and revision from Full S3TA. The same judge and criteria evaluate target-model responses across configurations.

Ablation Results. Original, Decompose, and Text-only succeed on exactly the same requests, yielding identical category-wise ASRs and an overall ASR of 6.84% (Table 4). Neither decomposition alone nor the constructed textual questions introduce additional successful attacks. As shown in Figure 8, subtask titles alone elicit little detail (D=1.79), while questions without their scene images lack the context needed to recover the intended task (A=2.07, D=2.14). Adding images in No-gate raises ASR to 87.97%, showing the substantial contribution of scene images within the S3TA pipeline: visible entities, relationships, and activity progress ground the assistance implied by the question. Checking and revision further raise ASR to 94.81%, with high alignment and detail (A=4.92, D=4.75), consistent with more recognizable scene cues improving task interpretation. This refinement is especially beneficial for Physical Harm, where ASR rises from 79.86% to 90.97%.

CategoryOriginalDecomposeText-onlyNo-gateFull S3TA
Illegal Activity0.000.000.0087.6395.88
Malware Generation11.3611.3611.3693.1895.45
Physical Harm10.4210.4210.4279.8690.97
Privacy Violence6.476.476.4794.9697.84
Overall6.846.846.8487.9794.81

Table 4. Ablation ASR (%) on Qwen3-VL-8B. Original uses unmodified MM-SafetyBench questions as text-only inputs. Bold marks row maxima.

Figure 8. Overall H/A/D scores for the ablation configurations.

Human Validation of the Judge

To assess the reliability of our automated evaluation, we compare GPT-4o-mini’s judgments with independent human annotations on 50 InternVL3.5-8B responses. Three annotators independently evaluate each response using the same H/A/D rating rubrics, given only the original harmful request and model response, without access to automated judgments. We derive binary success labels using the same evaluation criterion and define the human reference by majority vote. Table 5 reports mean rubric scores and ASRs for the judge, individual annotators, and human aggregate. GPT-4o-mini achieves 92% agreement with the human majority (Cohen’s κ=0.78\kappa= 0.78 [7]), while pairwise human agreement ranges from 90% to 94%, with Fleiss’ κ=0.75\kappa= 0.75 [9] across all three annotators. Individual annotators yield ASRs of 76–82%, while the human majority and automated judge yield 80% and 72%, respectively. These results indicate high agreement between automated and human judgments of attack success on the evaluated sample.

EvaluatorH (1–10)A (1–5)D (1–5)ASR (%)
Human 15.784.123.8276
Human 25.103.863.9482
Human 35.544.123.9082
Human aggregate5.474.033.8980
GPT-4o-mini6.424.103.6272

Table 5. Human and automated evaluation of 50 InternVL3.5-8B responses. Human aggregate uses mean H/A/D scores across annotators and majority-vote labels for ASR.

S3TA Against Existing Defenses

Text-side defenses. We evaluate text-side input defenses on the MM-SafetyBench requests used to evaluate S3TA on Qwen3-VL-8B, including Llama Guard 3 [14], NeMo perplexity heuristics [26], and GPT-2 perplexity filters with p95 and p99 thresholds calibrated on 520 benign questions [36]. Each defense inspects the textual question accompanying a scene image. We block a request if any stage is flagged and compute the resulting ASR from the existing model responses. Neither Llama Guard 3 nor the combined NeMo heuristics flags any evaluated input, leaving ASR unchanged at 87.25%. PPL-p95 and PPL-p99 intercept 96 and 95 requests, respectively, but both retain an ASR of 81.00%, a reduction of only 6.25 percentage points. These results show that S3TA remains effective under both semantic moderation and perplexity-based text filtering.

Visual-side defenses. ShieldGemma 2-4B [39] is a Gemma 3-based model trained to classify images for safety policy violations. We apply it to the corresponding scene images using its built-in dangerous-content, sexually explicit-content, and violence/gore policies, blocking an image if any policy receives a violation probability of at least 0.5. The overall image interception rate is 21.81%, leaving 78.19% of scene images unblocked. Interception varies across categories, reaching 48.34% for Malware Generation but falling to 13.65% for Economic Harm and 18.02% for Privacy Violence. The remaining categories have interception rates between 20.63% and 25.06%. Visual moderation therefore detects more inputs than text-side filtering, but still leaves most scene-state images unblocked.

Together, these results show that screening text or images in isolation provides limited protection against S3TA. The safety failure arises as the VLM interprets the scene state and reasons toward harmful assistance, exposing a gap between detecting unsafe input content and safeguarding scene-grounded reasoning.

Conclusion

We identify scene states as a source of safety failures in vision-language models: interpreting a depicted event can lead a model to infer harmful assistance without an explicit request for the underlying objective. SceneSafetyBench establishes the prevalence of these failures across scene-state categories, visual styles, and eight VLMs. Comparative mechanistic analysis across three models shows that scene-state inputs resemble QR-related images in their overall safety-related profiles, while differing in signal strength, generation dynamics, and local component effects. Building on this formulation, S3TA constructs task-relevant scene states and achieves strong performance across open-weight and commercial VLMs. Its effectiveness under text-side filtering, together with the limited interception of scene images by visual moderation, highlights a gap in existing safeguards. Multimodal safety must address harmful assistance inferred from visual events, extending protection from input screening to scene-grounded reasoning.

LLM Usage Statement

We used deepseek-flash to plan scene states and generate and revise scene descriptions for S3TA, and Qwen-Image to render these descriptions into images. We used GPT-Image-2 to produce the stylized variants in SceneSafetyBench. GPT-4o-mini evaluated model responses using the harmfulness, intent alignment, and detail rubrics described in the paper. The target VLMs served as experimental subjects, while Qwen3-VL-8B provided image-checking feedback for scene revision within S3TA. We also used Codex and DeepSeek Harness to assist with manuscript drafting and revision, code development, and figure preparation.

References

  1. [1]Anthropic. Claude Haiku 4.5 system card. https://assets.anthropic.com/m/91928dd00d9bcb/original/Claude-Haiku-4-5-System-Card.pdf, 2025. 3
  2. [2]Aharon Azulay, Jan Dubiński, Zhuoyun Li, Atharv Mittal, and Yossi Gandelsman. Jailbreaking vision-language models through the visual modality. In Proceedings of the International Conference on Machine Learning, pages 4636–4670, 2026. 1, 2arxiv.org/abs/2605.00583
  3. [3]Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chუნjiang Ge, et al. Qwen3-VL technical report. arXiv preprint arXiv:2511.21631, 2025. 3, 4, 6arxiv.org/abs/2511.21631
  4. [4]Davide Caffagni, Federico Cocchi, Luca Barsellotti, Nicholas Moretalli, Sara Sarto, Lorenzo Baraldi, Lorenzo Baraldi, Marcella Cornia, and Rita Cucchiara. The revolution of multimodal large language models: A survey. In Findings of the Association for Computational Linguistics: ACL 2024, pages 13590–13618, 2024. 1
  5. [5]Renmiao Chen, Shiyao Cui, Xuancheng Huang, Chengwei Pan, Victor Shea-Jay Huang, QingLin Zhang, Xuan Ouyang, Zhexin Zhang, Hongning Wang, and Minlie Huang. JPS: Jailbreak multimodal large language models with collaborative visual perturbation and textual steering. In Proceedings of the 33rd ACM International Conference on Multimedia, pages 11756–11765, 2025. 1, 2DOI
  6. [6]Renmiao Chen, Yida Lu, Shiyao Cui, Xuan Ouyang, Victor Shea-Jay Huang, Shumin Zhang, Chengwei Pan, Han Qiu, and Minlie Huang. The side effects of being smart: Safety risks in MLLMs’ multi-image reasoning. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 36866–36882, 2026. 1
  7. [7]Jacob Cohen. A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20(1):37–46, 1960. 8DOI
  8. [8]DeepSeek-AI, Anyi Xu, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, et al. DeepSeek-V4: Towards highly efficient million-token context intelligence. arXiv preprint arXiv:2606.19348, 2026. 6arxiv.org/abs/2606.19348
  9. [9]Joseph L. Fleiss. Measuring nominal scale agreement among many raters. Psychological Bulletin, 76(5):378–382, 1971. 8DOI
  10. [10]Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, et al. Gemma 3 technical report. arXiv preprint arXiv:2503.19786, 2025. 3
  11. [11]GLM-V Team, Wenyi Hong, Wenmeng Yu, Xiaotao Gu, Guo Wang, Guobing Gan, Haomiao Tang, Jiale Cheng, Ji Qi, Junhui Ji, et al. GLM-V-4.5V and GLM-4.1V-Thinking: Towards versatile multimodal reasoning with scalable reinforcement learning. arXiv preprint arXiv:2507.01006, 2025. 3
  12. [12]Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. FigStep: Jailbreaking large vision-language models via typographic visual prompts. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 23951–23959, 2025. 1, 2, 4, 7arxiv.org/abs/2311.05608
  13. [13]Google DeepMind. Gemini 3.5 Flash-Lite. https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/, 2026. 3
  14. [14]Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama Guard: LLM-based input-output safeguard for human-AI conversations. arXiv preprint arXiv:2312.06674, 2023. 8arxiv.org/abs/2312.06674
  15. [15]Xiaojun Jia, Jie Liao, Qi Guo, Teng Ma, Simeng Qin, Ranjie Duan, Tianlin Li, Yihao Huang, Zhitao Zeng, Dongxian Wu, et al. OmniSafeBench-MM: A unified benchmark and toolbox for multimodal jailbreak attack-defense evaluation. arXiv preprint arXiv:2512.06589, 2025. 2, 3arxiv.org/abs/2512.06589
  16. [16]Yilei Jiang, Xinyan Gao, Tian shuo Peng, Yingshui Tan, Xiaoyong Zhu, Bo Zheng, and Xiangyu Yue. HiddenDetect: Detecting jailbreak attacks against multimodal large language models via monitoring hidden states. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14880–14893, 2025. 2, 4DOI
  17. [17]Pengcheng Li, Jie Zhang, Tianwei Zhang, Han Qiu, Zhang Kejun, Weiming Zhang, Nenghai Yu, and Wenbo Zhou. State-dependent safety failures in multi-turn language model interaction. In Proceedings of the International Conference on Machine Learning, pages 72598–72615, 2026. 1, 3arxiv.org/abs/2603.15684
  18. [18]Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and Ji-Rong Wen. Images are Achilles’ heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. In European Conference on Computer Vision, pages 174–189, 2024. 2, 3arxiv.org/abs/2403.09792
  19. [19]Haotian Liu, Chunyuан Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296–26306, 2024. 1, 4arxiv.org/abs/2310.03744
  20. [20]Junliang Liu, Jingyu Xiao, Wenxin Tang, Zhixian Wang, Zeping Xie, Wenxuan Wang, Minrun Zhang, and Shuangheng Yu. Benchmarking MLLM-based web understanding: Reasoning, robustness and safety. In Proceedings of the 2026 International Conference on Multimodal Retrieval, pages 2182–2190, 2026. 1arxiv.org/abs/2509.21782
  21. [21]Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. MM-SafetyBench: A benchmark for safety evaluation of multimodal large language models. In European Conference on Computer Vision, pages 386–403, 2024. 1, 2, 3, 4, 7arxiv.org/abs/2311.17600
  22. [22]Bingjun Guo, Jialin Guo, Yue Yao, and Xinpeng Ding. Adversarial style optimization: Enhancing VLM jailbreaks by GRPO-based stylistic triggers optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11–19, 2026. 1, 2arxiv.org/abs/2607.21619
  23. [23]Weidi Luo, Siyuan Ma, Xiaogeng Liu, Xiaoyu Guo, and Chaowei Xiao. JailBreakV: A benchmark for assessing the robustness of multimodal large language models against jailbreak attacks. In First Conference on Language Modeling, 2024. 3arxiv.org/abs/2404.03027
  24. [24]Teng Ma, Xiaojun Jia, Ranjie Duan, Xinfeng Li, Yihao Huang, Xiaoshuang Jia, Zhixuan Chu, and Wenqi Ren. Heuristic-induced multimodal risk distribution jailbreak attack for multimodal large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2686–2696, 2025. 1, 2
  25. [25]Ziqi Miao, Yi Ding, Lijun Li, and Jing Shao. Visual contextual attack: Jailbreaking MLLMs with image-driven context injection. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 9627–9644, 2025. 1, 2
  26. [26]NVIDIA. Jailbreak protection. https://docs.nvidia.com/nemo/guardrails/configure-guardrails/guardrail-catalog/jailbreak-protection. 8
  27. [27]OpenAI. GPT-4o mini: Advancing cost-efficient intelligence. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/, 2024. 3
  28. [28]OpenAI. GPT-5.6 Luna. https://developers.openai.com/api/docs/models/gpt-5.6-luna, 2026. 3
  29. [29]Chu-Jun Peng, Zhong-Wei Lin, Yu-Wen Tseng, Hong-Han Shuai, and Wen-Huang Cheng. State2State: Revealing compositional safety failures through state-transition reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 51106–51121, 2026. 1, 3DOI
  30. [30]Rylan Schaeffer, Dan Valentine, Luke Bailey, James Chua, Cristobal Eyzaguirre, Zane Durante, Joe Benton, Brando Miranda, Henry Sleight, Tony Wang, et al. Failures to find transferable image jailbreaks between vision-language models. In International Conference on Learning Representations, 2025. 2, 4arxiv.org/abs/2407.15211
  31. [31]Bingrui Sima, Linhua Cong, Wenxuan Wang, and Kun He. VisCRA: A visual chain reasoning attack for jailbreaking multimodal large language models. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 6131–6144, 2025. 1, 7arxiv.org/abs/2505.19684
  32. [32]Jiaxin Song, Yixu Wang, Jie Li, Xuan Tong, Rui Yu, Yan Teng, Xingjun Ma, and Yingchun Wang. JailBound: Jailbreaking internal safety boundaries of vision-language models. In Advances in Neural Information Processing Systems, pages 25943–25968, 2025. 1, 2arxiv.org/abs/2505.19610
  33. [33]Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, et al. InternVL3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265, 2025. 3, 4
  34. [34]Yu Wang, Xiaofei Zhou, Yichen Wang, Geyuan Zhang, and Tianxing He. Jailbreak large vision-language models through multi-modal linkage. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1466–1494, 2025. 1, 2, 4, 7arxiv.org/abs/2412.00473
  35. [35]Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-Image technical report. arXiv preprint arXiv:2508.02324, 2025. 6
  36. [36]Tongxi Wu, Jian Zhang, and Yang Gao. Furina: Fragmented uncertainty-driven refusal instability attack. In Proceedings of the International Conference on Machine Learning, pages 137443–137474, 2026. 8arxiv.org/abs/2605.26158
  37. [37]Yuan Xiong, Ziqi Miao, Lijun Li, Chen Qian, Jie Li, and Jing Shao. Contextual image attack: How visual context exposes multimodal safety vulnerabilities. In European Conference on Computer Vision, pages 240–257, 2026. 1, 2, 4, 7arxiv.org/abs/2512.02973
  38. [38]Tianyu Yu, Zefan Wang, Chongyi Wang, Fuwei Huang, Wenshuo Ma, Zhihui He, Tianchi Cai, Weize Chen, Yuxiang Huang, Ranchi Zhao, et al. MiniCPM-V 4.5: Cooking efficient MLLMs via architecture, data, and training recipe. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11704–11715, 2026. 3arxiv.org/abs/2509.18154
  39. [39]Wenjun Zeng, Dana Kurniawan, Ryan Mullins, Yuchi Liu, Tamogna Saha, Dirichi Ike-Njoku, Jindong Gu, Yiwen Song, Cai Xu, Jingjing Zhou, et al. ShieldGemma 2: Robust and tractable image content moderation. arXiv preprint arXiv:2504.01081, 2025. 8arxiv.org/abs/2504.01081
  40. [40]Yunzhi Zhang, Zizhang Li, Matt Zhou, Shangzhe Wu, and Jiajun Wu. The Scene Language: Representing scenes with programs, words, and embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24625–24634, 2025. 1arxiv.org/abs/2410.16770
  41. [41]Ziyi Zhang, Zhen Sun, Zongmin Zhang, Jihui Guo, and Xinlei He. FC-Attack: Jailbreaking multimodal large language models via auto-generated flowcharts. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 9299–9316, 2025. 1, 2, 7arxiv.org/abs/2502.21059
  42. [42]Shiji Zhao, Shukun Xiong, Yao Huang, Jin Yan, Zhenyu Wu, Jiyang Guan, Ranjie Duan, Jialing Tao, Hui Xue, and Xingying Wei. VRSA: Jailbreaking multimodal large language models through visual reasoning sequential attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9412–9421, 2026. 1, 2arxiv.org/abs/2512.05853
  43. [43]Kaiwen Zhou, Chengzhi Liu, Xuandong Zhao, Anderson Compalas, Dawn Song, and Xin Eric Wang. Multimodal situational safety. In International Conference on Learning Representations, 2025. 1, 3arxiv.org/abs/2410.06172

Paper details

vlnc:W_2hnsckm3
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Codex, DeepSeek Harness

View statement
Looking for related papers…
Contents