-
Notifications
You must be signed in to change notification settings - Fork 9
Expand file tree
/
Copy pathindex.html
More file actions
880 lines (791 loc) · 31.5 KB
/
Copy pathindex.html
File metadata and controls
880 lines (791 loc) · 31.5 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>MM-Plan: Visual Exclusivity Attacks via Agentic Planning</title>
<meta name="description" content="MM-Plan: Automatic Multimodal Red Teaming via Agentic Planning - A framework for evaluating multimodal AI safety through Visual Exclusivity attacks.">
<!-- Fonts -->
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link href="https://fonts.googleapis.com/css2?family=Noto+Sans:wght@400;500;600;700&family=Noto+Serif:wght@400;500;600&display=swap" rel="stylesheet">
<!-- Font Awesome for icons -->
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.4.0/css/all.min.css">
<style>
:root {
--primary-color: #059669;
--primary-dark: #047857;
--secondary-color: #10b981;
--accent-color: #34d399;
--text-color: #1f2937;
--text-light: #6b7280;
--bg-color: #ffffff;
--bg-secondary: #f0fdf4;
--border-color: #d1fae5;
--success-color: #10b981;
--warning-color: #ef4444;
}
* {
margin: 0;
padding: 0;
box-sizing: border-box;
}
body {
font-family: 'Noto Sans', -apple-system, BlinkMacSystemFont, sans-serif;
line-height: 1.7;
color: var(--text-color);
background-color: var(--bg-color);
}
/* Hero Section */
.hero {
background: linear-gradient(135deg, #064e3b 0%, #047857 50%, #059669 100%);
color: white;
padding: 80px 20px 60px;
text-align: center;
}
.hero h1 {
font-family: 'Noto Serif', Georgia, serif;
font-size: clamp(2rem, 5vw, 3.2rem);
font-weight: 600;
margin-bottom: 20px;
line-height: 1.2;
}
.hero .subtitle {
font-size: clamp(1.1rem, 2.5vw, 1.4rem);
opacity: 0.95;
max-width: 800px;
margin: 0 auto 30px;
font-weight: 400;
}
.venue-badge {
display: inline-block;
background: linear-gradient(135deg, rgba(255, 215, 0, 0.25), rgba(255, 165, 0, 0.2));
padding: 8px 20px;
border-radius: 30px;
font-size: 0.95rem;
margin-bottom: 30px;
border: 1px solid rgba(255, 215, 0, 0.6);
font-weight: 600;
}
/* Authors */
.authors {
display: flex;
flex-wrap: wrap;
justify-content: center;
gap: 8px 25px;
margin-bottom: 15px;
font-size: 1.05rem;
}
.authors a {
color: white;
text-decoration: none;
transition: opacity 0.2s;
}
.authors a:hover {
opacity: 0.8;
text-decoration: underline;
}
.authors sup {
font-size: 0.7em;
margin-left: 2px;
}
.affiliations {
font-size: 0.95rem;
opacity: 0.9;
margin-bottom: 30px;
}
/* Buttons */
.buttons {
display: flex;
flex-wrap: wrap;
justify-content: center;
gap: 12px;
margin-top: 30px;
}
.btn {
display: inline-flex;
align-items: center;
gap: 8px;
padding: 12px 24px;
border-radius: 8px;
text-decoration: none;
font-weight: 500;
font-size: 0.95rem;
transition: all 0.2s;
}
.btn-primary {
background: white;
color: var(--primary-dark);
}
.btn-primary:hover {
background: #f0f0f0;
transform: translateY(-2px);
}
.btn-secondary {
background: rgba(255,255,255,0.15);
color: white;
border: 1px solid rgba(255,255,255,0.3);
}
.btn-secondary:hover {
background: rgba(255,255,255,0.25);
transform: translateY(-2px);
}
/* Container */
.container {
max-width: 1000px;
margin: 0 auto;
padding: 0 20px;
}
/* Sections */
section {
padding: 60px 0;
}
section:nth-child(even) {
background: var(--bg-secondary);
}
h2 {
font-family: 'Noto Serif', Georgia, serif;
font-size: 1.8rem;
font-weight: 600;
margin-bottom: 25px;
color: var(--text-color);
}
h3 {
font-size: 1.3rem;
font-weight: 600;
margin: 30px 0 15px;
color: var(--text-color);
}
p {
margin-bottom: 15px;
color: var(--text-color);
}
/* Warning Banner */
.warning-banner {
background: #fef2f2;
border: 1px solid #fecaca;
border-left: 4px solid #ef4444;
padding: 15px 20px;
border-radius: 0 8px 8px 0;
margin-bottom: 30px;
}
.warning-banner strong {
color: #dc2626;
}
/* Abstract */
.abstract {
background: white;
border: 1px solid var(--border-color);
border-radius: 12px;
padding: 30px;
font-size: 1.05rem;
line-height: 1.8;
}
/* Figure */
.figure {
margin: 40px 0;
text-align: center;
}
.figure img {
max-width: 100%;
border-radius: 8px;
box-shadow: 0 4px 20px rgba(0,0,0,0.1);
}
.figure-caption {
margin-top: 15px;
font-size: 0.95rem;
color: var(--text-light);
max-width: 800px;
margin-left: auto;
margin-right: auto;
}
/* Tables */
.table-container {
overflow-x: auto;
margin: 30px 0;
}
table {
width: 100%;
border-collapse: collapse;
font-size: 0.9rem;
background: white;
border-radius: 8px;
overflow: hidden;
box-shadow: 0 2px 10px rgba(0,0,0,0.05);
}
th, td {
padding: 12px 15px;
text-align: center;
border-bottom: 1px solid var(--border-color);
}
th {
background: #f8fafc;
font-weight: 600;
color: var(--text-color);
}
thead tr:first-child th {
border-bottom: 2px solid var(--border-color);
}
td:first-child, th:first-child {
text-align: left;
font-weight: 500;
}
.highlight-row {
background: #ecfdf5 !important;
}
.highlight-row td {
font-weight: 600;
color: var(--primary-dark);
}
.best-score {
color: var(--primary-dark);
font-weight: 700;
}
/* Key Results Grid */
.results-grid {
display: grid;
grid-template-columns: repeat(auto-fit, minmax(200px, 1fr));
gap: 20px;
margin: 30px 0;
}
.result-card {
background: white;
border: 1px solid var(--border-color);
border-radius: 12px;
padding: 25px;
text-align: center;
}
.result-card .number {
font-size: 2.5rem;
font-weight: 700;
color: var(--primary-color);
line-height: 1;
}
.result-card .label {
margin-top: 8px;
font-size: 0.95rem;
color: var(--text-light);
}
.result-card .sublabel {
font-size: 0.85rem;
color: var(--text-light);
margin-top: 4px;
}
/* Contributions List */
.contributions-list {
list-style: none;
padding: 0;
}
.contributions-list li {
display: flex;
align-items: flex-start;
gap: 15px;
margin-bottom: 20px;
padding: 20px;
background: white;
border-radius: 10px;
border: 1px solid var(--border-color);
}
.contributions-list .icon {
width: 40px;
height: 40px;
background: linear-gradient(135deg, var(--primary-color), var(--secondary-color));
border-radius: 10px;
display: flex;
align-items: center;
justify-content: center;
color: white;
flex-shrink: 0;
}
.contributions-list h4 {
font-weight: 600;
margin-bottom: 5px;
}
.contributions-list p {
margin: 0;
color: var(--text-light);
font-size: 0.95rem;
}
/* Comparison boxes */
.comparison-grid {
display: grid;
grid-template-columns: repeat(auto-fit, minmax(300px, 1fr));
gap: 25px;
margin: 30px 0;
}
.comparison-box {
background: white;
border: 1px solid var(--border-color);
border-radius: 12px;
padding: 25px;
}
.comparison-box h4 {
font-size: 1.1rem;
margin-bottom: 12px;
display: flex;
align-items: center;
gap: 10px;
}
.comparison-box.prior h4 {
color: var(--text-light);
}
.comparison-box.ours h4 {
color: var(--primary-color);
}
.comparison-box ul {
margin-left: 20px;
color: var(--text-light);
}
.comparison-box ul li {
margin-bottom: 8px;
}
/* BibTeX */
.bibtex-container {
background: #1e293b;
border-radius: 10px;
padding: 20px;
position: relative;
margin-top: 20px;
}
.bibtex-container pre {
color: #e2e8f0;
font-family: 'Consolas', 'Monaco', monospace;
font-size: 0.85rem;
overflow-x: auto;
white-space: pre-wrap;
word-wrap: break-word;
}
.copy-btn {
position: absolute;
top: 10px;
right: 10px;
background: rgba(255,255,255,0.1);
border: none;
color: white;
padding: 8px 12px;
border-radius: 6px;
cursor: pointer;
font-size: 0.85rem;
transition: background 0.2s;
}
.copy-btn:hover {
background: rgba(255,255,255,0.2);
}
/* Footer */
footer {
background: #1e293b;
color: #94a3b8;
padding: 40px 20px;
text-align: center;
}
footer a {
color: #6ee7b7;
text-decoration: none;
}
footer a:hover {
text-decoration: underline;
}
/* Responsive */
@media (max-width: 768px) {
.hero {
padding: 60px 20px 40px;
}
.authors {
font-size: 0.95rem;
}
section {
padding: 40px 0;
}
h2 {
font-size: 1.5rem;
}
.buttons {
flex-direction: column;
align-items: center;
}
.btn {
width: 100%;
max-width: 250px;
justify-content: center;
}
}
</style>
</head>
<body>
<!-- Hero Section -->
<header class="hero">
<div class="container">
<div class="venue-badge">🏆 Best Short Paper Award — ICLR 2026 Workshop AIWILD</div>
<h1>Visual Exclusivity Attacks: Automatic Multimodal Red Teaming via Agentic Planning</h1>
<p class="subtitle">MM-Plan: A framework that reformulates multimodal jailbreaking as agentic planning, achieving state-of-the-art attack success rates against frontier MLLMs</p>
<div class="authors">
<span><a href="https://yunbeizhang.github.io/">Yunbei Zhang</a><sup>1*</sup></span>
<span>Yingqiang Ge<sup>2</sup></span>
<span>Weijie Xu<sup>2</sup></span>
<span>Yuhui Xu<sup>2</sup></span>
<span>Jihun Hamm<sup>1</sup></span>
<span>Chandan K. Reddy<sup>2</sup></span>
</div>
<p class="affiliations"><sup>1</sup>Tulane University <sup>2</sup>Amazon</p>
<p class="affiliations" style="font-size: 0.85rem; opacity: 0.8; margin-top: 8px;">*Work done during an internship at Amazon</p>
<div class="buttons">
<a href="https://arxiv.org/abs/2603.20198" class="btn btn-secondary">
<i class="fas fa-file-pdf"></i> arXiv
</a>
<a href="https://github.com/yunbeizhang/MM-Plan" class="btn btn-secondary">
<i class="fab fa-github"></i> Code
</a>
<a href="https://huggingface.co/datasets/zybeich/VE-Safety" class="btn btn-secondary">
<i class="fas fa-database"></i> VE-Safety Dataset
</a>
</div>
</div>
</header>
<!-- Abstract Section -->
<section>
<div class="container">
<h2>Abstract</h2>
<div class="warning-banner">
<strong>Warning:</strong> This paper contains examples of potentially harmful content for research purposes.
</div>
<div class="abstract">
<p>Current multimodal red teaming treats images as wrappers for malicious payloads via typography or adversarial noise. These attacks are structurally brittle, as standard defenses neutralize them once the payload is exposed.</p>
<p>We introduce <strong>Visual Exclusivity (VE)</strong>, a more resilient <em>Image-as-Basis</em> threat where harm emerges only through reasoning over visual content such as technical schematics. To systematically exploit VE, we propose <strong>Multimodal Multi-turn Agentic Planning (MM-Plan)</strong>, a framework that reframes jailbreaking from turn-by-turn reaction to global plan synthesis.</p>
<p>MM-Plan trains an attacker planner to synthesize comprehensive, multi-turn strategies, optimized via Group Relative Policy Optimization (GRPO), enabling self-discovery of effective strategies without human supervision. To rigorously benchmark this reasoning-dependent threat, we introduce <strong>VE-Safety</strong>, a human-curated dataset filling a critical gap in evaluating high-risk technical visual understanding.</p>
<p>MM-Plan achieves <strong>46.3% attack success rate against Claude 4.5 Sonnet</strong> and <strong>13.8% against GPT-5</strong>, outperforming baselines by 2–5× where existing methods largely fail. These findings reveal that frontier models remain vulnerable to agentic multimodal attacks, exposing a critical gap in current safety alignment.</p>
</div>
</div>
</section>
<!-- Key Results Section -->
<section>
<div class="container">
<h2>Key Results</h2>
<div class="results-grid">
<div class="result-card">
<div class="number">46.3%</div>
<div class="label">ASR on Claude 4.5 Sonnet</div>
<div class="sublabel">vs 24.4% best baseline</div>
</div>
<div class="result-card">
<div class="number">13.8%</div>
<div class="label">ASR on GPT-5</div>
<div class="sublabel">vs 3.1% best baseline</div>
</div>
<div class="result-card">
<div class="number">2–5×</div>
<div class="label">Improvement</div>
<div class="sublabel">over prior methods</div>
</div>
<div class="result-card">
<div class="number">440</div>
<div class="label">VE-Safety Instances</div>
<div class="sublabel">15 safety categories</div>
</div>
</div>
</div>
</section>
<!-- Method Overview -->
<section>
<div class="container">
<h2>Method Overview</h2>
<div class="figure">
<img src="static/mm_plan_workflow.png" alt="MM-Plan Framework Overview">
<p class="figure-caption"><strong>Figure:</strong> MM-Plan framework. Given a harmful goal and image, our Attacker Planner generates complete multi-turn strategies in a single pass. Plans are sampled and executed against victim MLLMs, with rewards collected from a judge model. The policy is updated via GRPO based on relative plan performance.</p>
</div>
<h3>Why Agentic Planning?</h3>
<div class="comparison-grid">
<div class="comparison-box prior">
<h4><i class="fas fa-times-circle"></i> Prior Approaches</h4>
<ul>
<li><strong>Sequential RL:</strong> Suffers from myopia, optimizing for immediate rewards</li>
<li><strong>Iterative Search:</strong> Scales poorly with K<sup>N</sup> trajectories for N-turn dialogues</li>
<li><strong>Wrapper-based Attacks:</strong> Easily neutralized by OCR-aware filters</li>
</ul>
</div>
<div class="comparison-box ours">
<h4><i class="fas fa-check-circle"></i> MM-Plan (Ours)</h4>
<ul>
<li><strong>Global Planning:</strong> Synthesizes complete strategy in one pass</li>
<li><strong>Linear Scaling:</strong> Only K × N steps for K sampled plans</li>
<li><strong>Visual Operations:</strong> Exploits image reasoning, not text wrappers</li>
</ul>
</div>
</div>
</div>
</section>
<!-- Visual Exclusivity Section -->
<section>
<div class="container">
<h2>Visual Exclusivity: A New Threat Model</h2>
<div class="figure">
<img src="static/motivation_ve.png" alt="Image-as-Wrapper vs Image-as-Basis" style="max-width: 700px;">
<p class="figure-caption"><strong>Image-as-Wrapper vs. Image-as-Basis.</strong> Prior attacks (top) embed harmful instructions typographically within images. In contrast, Visual Exclusivity (bottom) presents an <em>Image-as-Basis</em> threat where text input alone is insufficient—the harmful goal requires reasoning about spatial and functional relationships exclusive to the image.</p>
</div>
<p>Unlike prior "wrapper-based" attacks where images merely conceal text payloads, <strong>Visual Exclusivity (VE)</strong> exploits the model's own visual reasoning capabilities. In VE attacks:</p>
<ul style="margin-left: 25px; margin-bottom: 20px;">
<li>The text query appears innocuous (e.g., "How do I assemble this?")</li>
<li>The image contains no adversarial noise or hidden typography</li>
<li>Harm materializes only when the model correctly interprets spatial/functional relationships in the image</li>
</ul>
<p>This dependency renders standard defenses largely ineffective: OCR cannot extract payloads that don't exist in text form, and caption-based screening cannot capture precise structural details required for harm.</p>
</div>
</section>
<!-- Main Results Table -->
<section>
<div class="container">
<h2>Main Results on VE-Safety</h2>
<p>Attack Success Rate (ASR %) across 8 frontier MLLMs. MM-Plan significantly outperforms all baselines, especially on heavily defended proprietary models.</p>
<div class="table-container">
<table>
<thead>
<tr>
<th rowspan="2">Method</th>
<th colspan="3">Open-Weight</th>
<th colspan="5">Proprietary</th>
</tr>
<tr>
<th>Llama-3.2-11B</th>
<th>InternVL3-8B</th>
<th>Qwen3-VL-8B</th>
<th>GPT-4o</th>
<th>GPT-5</th>
<th>Sonnet 3.7</th>
<th>Sonnet 4.5</th>
<th>Gemini 2.5 Pro</th>
</tr>
</thead>
<tbody>
<tr>
<td>Direct Request</td>
<td>13.4</td>
<td>27.2</td>
<td>11.9</td>
<td>5.0</td>
<td>0.6</td>
<td>4.7</td>
<td>8.4</td>
<td>9.7</td>
</tr>
<tr>
<td>Direct Plan</td>
<td>18.1</td>
<td>34.7</td>
<td>22.5</td>
<td>9.4</td>
<td>0.9</td>
<td>8.1</td>
<td>9.7</td>
<td>11.9</td>
</tr>
<tr>
<td>FigStep</td>
<td>23.8</td>
<td>44.4</td>
<td>33.1</td>
<td>6.6</td>
<td>0.6</td>
<td>13.4</td>
<td>24.4</td>
<td>11.3</td>
</tr>
<tr>
<td>SI-Attack</td>
<td>25.6</td>
<td>31.9</td>
<td>29.1</td>
<td>8.1</td>
<td>1.9</td>
<td>12.8</td>
<td>15.6</td>
<td>12.5</td>
</tr>
<tr>
<td>SSA</td>
<td>25.3</td>
<td>39.1</td>
<td>29.4</td>
<td>6.3</td>
<td>1.6</td>
<td>9.7</td>
<td>15.9</td>
<td>12.2</td>
</tr>
<tr>
<td>Crescendo</td>
<td>21.9</td>
<td>45.0</td>
<td>33.8</td>
<td>14.4</td>
<td>3.1</td>
<td>15.0</td>
<td>18.1</td>
<td>15.9</td>
</tr>
<tr class="highlight-row">
<td><strong>MM-Plan (Ours)</strong></td>
<td class="best-score">64.4*</td>
<td class="best-score">65.0*</td>
<td class="best-score">54.4*</td>
<td class="best-score">36.9*</td>
<td class="best-score">13.8*</td>
<td class="best-score">27.2*</td>
<td class="best-score">46.3*</td>
<td class="best-score">43.8*</td>
</tr>
</tbody>
</table>
</div>
<p style="font-size: 0.9rem; color: var(--text-light); margin-top: 10px;">* Statistically significant improvement (p ≤ 0.05) over second-best method.</p>
</div>
</section>
<!-- VE-Safety Dataset -->
<section>
<div class="container">
<h2>VE-Safety Benchmark</h2>
<p>We introduce <strong>VE-Safety</strong>, the first benchmark specifically targeting the <em>Image-as-Basis</em> threat model with real-world technical imagery.</p>
<div class="results-grid" style="margin-top: 30px;">
<div class="result-card">
<div class="number">440</div>
<div class="label">Human-Curated Instances</div>
</div>
<div class="result-card">
<div class="number">15</div>
<div class="label">Safety Categories</div>
</div>
<div class="result-card">
<div class="number">100%</div>
<div class="label">Real-World Images</div>
</div>
</div>
<h3>Dataset Characteristics</h3>
<ul style="margin-left: 25px;">
<li><strong>Text Insufficiency:</strong> Harmful goals are unattainable via text alone</li>
<li><strong>Visual Sufficiency:</strong> Goals become achievable given the visual information</li>
<li><strong>Real-World Imagery:</strong> Technical schematics, circuit diagrams, floor plans, chemical formulas</li>
<li><strong>Multi-Turn Focus:</strong> Designed for evaluating multi-turn attack strategies</li>
</ul>
<div class="table-container" style="margin-top: 30px;">
<table>
<thead>
<tr>
<th>Benchmark</th>
<th>Human-Curated</th>
<th>Image Type</th>
<th>Visual Role</th>
<th>Multi-Turn</th>
</tr>
</thead>
<tbody>
<tr>
<td>FigStep</td>
<td>✗</td>
<td>Typographic</td>
<td>Image-as-Wrapper</td>
<td>✗</td>
</tr>
<tr>
<td>HADES</td>
<td>✗</td>
<td>Typo. / Adv. Noise</td>
<td>Image-as-Wrapper</td>
<td>✗</td>
</tr>
<tr>
<td>MM-SafetyBench</td>
<td>✗</td>
<td>Typo. / SD</td>
<td>Image-as-Wrapper</td>
<td>✗</td>
</tr>
<tr>
<td>HarmBench (MM)</td>
<td>✗</td>
<td>SD / Real</td>
<td>Image-as-Basis</td>
<td>✗</td>
</tr>
<tr class="highlight-row">
<td><strong>VE-Safety (Ours)</strong></td>
<td>✓</td>
<td><strong>Real</strong></td>
<td><strong>Image-as-Basis</strong></td>
<td>✓</td>
</tr>
</tbody>
</table>
</div>
</div>
</section>
<!-- Contributions -->
<section>
<div class="container">
<h2>Contributions</h2>
<ul class="contributions-list">
<li>
<div class="icon"><i class="fas fa-eye"></i></div>
<div>
<h4>Visual Exclusivity (VE)</h4>
<p>We formalize a new multimodal vulnerability where harmful goals require visual reasoning about image content, providing criteria that distinguish VE from wrapper-based attacks.</p>
</div>
</li>
<li>
<div class="icon"><i class="fas fa-database"></i></div>
<div>
<h4>VE-Safety Benchmark</h4>
<p>We construct the first benchmark targeting Image-as-Basis threats, comprising 440 human-curated instances across 15 safety categories with verified non-textual irreducibility.</p>
</div>
</li>
<li>
<div class="icon"><i class="fas fa-robot"></i></div>
<div>
<h4>MM-Plan Framework</h4>
<p>We propose a multimodal agentic planning framework that achieves 2–5× higher attack success rates than search-based and turn-by-turn baselines across frontier MLLMs.</p>
</div>
</li>
</ul>
</div>
</section>
<!-- Citation -->
<section>
<div class="container">
<h2>Citation</h2>
<p>If you find our work useful, please cite our paper:</p>
<div class="bibtex-container">
<button class="copy-btn" onclick="copyBibtex()">
<i class="fas fa-copy"></i> Copy
</button>
<pre><code>@article{zhang2026mmplan,
title={Visual Exclusivity Attacks: Automatic Multimodal Red Teaming via Agentic Planning},
author={Zhang, Yunbei and Ge, Yingqiang and Xu, Weijie and Xu, Yuhui and Hamm, Jihun and Reddy, Chandan K.},
journal={arXiv preprint arXiv:2603.20198},
year={2026}
}</code></pre>
</div>
</div>
</section>
<!-- Footer -->
<footer>
<div class="container">
<p>MM-Plan © 2026 | <a href="https://github.com/yunbeizhang/MM-Plan">GitHub</a> | <a href="mailto:yzhang111@tulane.edu">Contact</a></p>
<p style="margin-top: 10px; font-size: 0.85rem;">Website template inspired by <a href="https://nerfies.github.io/">Nerfies</a> and academic project pages.</p>
</div>
</footer>
<script>
function copyBibtex() {
const bibtex = `@article{zhang2026mmplan,
title={Visual Exclusivity Attacks: Automatic Multimodal Red Teaming via Agentic Planning},
author={Zhang, Yunbei and Ge, Yingqiang and Xu, Weijie and Xu, Yuhui and Hamm, Jihun and Reddy, Chandan K.},
journal={arXiv preprint arXiv:2603.20198},
year={2026}
}`;
navigator.clipboard.writeText(bibtex).then(() => {
const btn = document.querySelector('.copy-btn');
btn.innerHTML = '<i class="fas fa-check"></i> Copied!';
setTimeout(() => {
btn.innerHTML = '<i class="fas fa-copy"></i> Copy';
}, 2000);
});
}
</script>
</body>
</html>