-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
2493 lines (2402 loc) · 124 KB
/
index.html
File metadata and controls
2493 lines (2402 loc) · 124 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
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="description"
content="RiOSWorld: Benchmarking the Risk of Multimodal Computer-Use Agents">
<meta name="keywords" content="General-Purpose Agent, Human Computer Interaction">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>RiOSWorld: Benchmarking the Risk of Multimodal Computer-Use Agents</title>
<script type="module" src="https://md-block.verou.me/md-block.js"></script>
<!-- Global site tag (gtag.js) - Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=G-PYVRSFMDRL"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag() {
dataLayer.push(arguments);
}
gtag('js', new Date());
gtag('config', 'G-PYVRSFMDRL');
</script>
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
<link rel="stylesheet" href="./static/css/bulma.min.css">
<link rel="stylesheet" href="./static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="./static/css/bulma-slider.min.css">
<link rel="stylesheet" href="./static/css/fontawesome.all.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="./static/css/index.css">
<link rel="icon" href="static/images/demon_logo.png">
<link rel="stylesheet" href="./stylesheets/layout.css">
<link rel="stylesheet" href="./stylesheets/index.css">
<link rel="stylesheet" href="./bowe_componets/css/bootstrap.table.min.css">
<style>
.first-rank-star {
color: gold;
font-size: 1.2em;
display: inline-block;
animation: pulse 1.5s infinite;
margin-left: 5px;
}
.first-rank-row {
background-color: rgba(255, 215, 0, 0.1) !important;
}
@keyframes pulse {
0% { transform: scale(1); }
50% { transform: scale(1.3); }
100% { transform: scale(1); }
}
.trajectory-link {
display: flex;
flex-direction: column;
align-items: center;
color: #3273dc;
text-decoration: none;
}
.trajectory-preview {
width: 60px;
height: 40px;
object-fit: cover;
margin-top: 5px;
border-radius: 3px;
transition: all 0.3s ease;
}
.trajectory-link:hover .trajectory-preview {
width: 80px;
height: 60px;
}
/* Experiments Table Styles */
.experiments-table {
font-size: 0.9em;
border-collapse: collapse;
width: 100%;
margin: 20px 0;
}
.experiments-table th, .experiments-table td {
border: 1px solid #ddd;
padding: 8px;
text-align: center;
}
.experiments-table th {
background-color: #f5f5f5;
font-weight: bold;
}
.experiments-table .model-name {
text-align: left;
font-weight: 500;
}
.experiments-table .average-row {
background-color: #f0f0f0;
font-weight: bold;
}
.risk-high { background-color: #ffebee; }
.risk-medium { background-color: #fff3e0; }
.risk-low { background-color: #e8f5e8; }
.table-container {
overflow-x: auto;
margin: 20px 0;
}
.instruction {
color: #6C8EBF;
font-family: monospace;
}
.config {
color: #60A917;
font-family: monospace;
}
.halfway-config {
color: #B85450;
font-family: monospace;
}
.evaluator {
color: #FF9933;
font-family: monospace;
}
.risk-evaluator {
color: #CCCC00;
font-family: monospace;
}
</style>
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="./static/js/fontawesome.all.min.js"></script>
<script src="./static/js/bulma-carousel.min.js"></script>
<script src="./static/js/bulma-slider.min.js"></script>
<script src="./static/js/index.js"></script>
</head>
<body>
<nav class="navbar" role="navigation" aria-label="main navigation">
<div class="navbar-brand">
<a role="button" class="navbar-burger" aria-label="menu" aria-expanded="false">
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
</a>
</div>
<div class="navbar-menu">
<div class="navbar-start" style="flex-grow: 1; justify-content: center;">
<a class="navbar-item" href="https://open-trust-lab.vercel.app/">
<span class="icon">
<i class="fas fa-home"></i>
</span>
</a>
<div class="navbar-item has-dropdown is-hoverable">
<a class="navbar-link">
More Research
</a>
<div class="navbar-dropdown">
<a class="navbar-item" href="https://os-world.github.io/">
OSWorld
</a>
<a class="navbar-item" href="https://github.com/SALT-NLP/PopupAttack">
PopupAttack
</a>
<a class="navbar-item" href="https://safearena.github.io/">
SafeArena
</a>
<a class="navbar-item" href="https://arxiv.org/pdf/2504.18575?">
WASP
</a>
<a class="navbar-item" href="https://github.com/xbmxb/EnvDistraction">
Env. Distraction
</a>
<!-- <a class="navbar-item" href="https://ds1000-code-gen.github.io/">
DS-1000
</a>
<a class="navbar-item" href="https://instructor-embedding.github.io/">
Instructor
</a>
<a class="navbar-item" href="https://text-to-reward.github.io/">
Text2Reward
</a>
<a class="navbar-item" href="https://github.com/xlang-ai/OpenAgents">
OpenAgents
</a>
<a class="navbar-item" href="https://github.com/OpenLemur/lemur">
Lemur-70B
</a>
<a class="navbar-item" href="https://arks-codegen.github.io/">
ARKS
</a>
<a class="navbar-item" href="https://brightbenchmark.github.io/">
BRIGHT
</a>
<a class="navbar-item" href="https://spider2-v.github.io/">
Spider2-V
</a> -->
</div>
</div>
</div>
</div>
</nav>
<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-1 publication-title">
RiOSWorld: Benchmarking the Risk of Multimodal Computer-Use Agents
</h1>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://scholar.google.com/citations?user=3NSeUiwAAAAJ&hl=en">Jingyi Yang</a><sup>2</sup>,</span>
<span class="author-block">
<a href="https://github.com/ShaoShuai0605">Shuai Shao</a><sup>3</sup>,</span>
<span class="author-block">
<a href="https://shenqildr.github.io/">Dongrui Liu</a><sup>1</sup>,</span>
<span class="author-block">
<a href="https://amandajshao.github.io/">Jing Shao</a><sup>1</sup>,</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>1</sup>Shanghai Artificial Intelligence Laboratory,</span>
<span class="author-block"><sup>2</sup>University of Science and Technology of China,</span>
<span class="author-block"><sup>3</sup>Shanghai Jiao Tong University</span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<span class="link-block">
<a href="https://arxiv.org/pdf/2506.00618" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>Paper</span>
</a>
</span>
<span class="link-block">
<a href="https://github.com/yjyddq/RiOSWorld"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Code</span>
</a>
</span>
<!-- Doc. -->
<!-- <span class="link-block">
<a href="https://timothyxxx.github.io/OSWorld/"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="far fa-bookmark"></i>
</span>
<span>Doc</span>
</a>
</span> -->
<!-- Dataset Link. -->
<span class="link-block">
<a href="https://github.com/yjyddq/RiOSWorld/tree/main/evaluation_risk_examples"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="far fa-images"></i>
</span>
<span>Data</span>
</a>
</span>
<span class="link-block">
<a href="https://huggingface.co/datasets/JY-Young/RiOSWorld"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<img src="static/images/hf-logo.png" alt="Huggingface logo" style="height: 1em; width: auto;">
</span>
<span>Huggingface</span>
</a>
</span>
<!-- Data Viewer. -->
<!-- <span class="link-block">
<a href="explorer.html" target="_blank" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fa fa-desktop"></i>
</span>
<span>Data Viewer</span>
</a>
</span> -->
<!-- Slides Link. -->
<!-- <span class="link-block">
<a href="https://docs.google.com/presentation/d/1-r889Nb9n7SeZqrj-ryNqJLoMzp7aGNU2ihO8nUdEcE/edit?usp=sharing"
target="_blank" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-google"></i>
</span>
<span>Slides</span>
</a>
</span> -->
<!-- Twitter Link. -->
<!-- <span class="link-block">
<a href="https://twitter.com/TianbaoX/status/1778781521253667267" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-twitter"></i>
</span>
<span>Twitter</span>
</a>
</span> -->
<!-- Discord Link. -->
<!-- <span class="link-block">
<a href="https://discord.gg/4Gnw7eTEZR" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-discord"></i>
</span>
<span>Discord</span>
</a>
</span> -->
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- <section class="section">
<div class="container is-max-desktop">
<div class="column is-full">
<video controls muted loop autoplay width="100%">
<source src="static/videos/main.mp4" type="video/mp4">
</video>
</div>
</div>
</section> -->
<section class="section">
<div class="container is-max-desktop">
<!-- Main Figure. -->
<h2 class="title is-3"></h2>
<div class="content has-text-justified">
<img src="static/images/benchmark.png" width="100%" alt="riosworld task_demonstration"
class="responsive-image">
<md-block>
**RiOSWorld** is a comprehensive, scalable, and real-world benchmark for evaluating the safety risk of computer-use agents based on Virtual Machine. It supports dynamic threat deployments (e.g., phishing email, pop-ups/Ads and reCAPTCHA, etc.).
**Upper section** indicates the risk category and illustrates several triggering risks, with environmental risks shown on the left and user-originated risks on the right.
**Lower section** depicts our evaluation pipeline. We employ: 1) a rule-based evaluator to evaluate the risk goal completion, and 2) an LLM-as-a-judge-based evaluator to evaluate the risk goal intention.
</md-block>
</div>
<!--/ Main Figure. -->
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Video Demos. -->
<h2 class="title is-3">Video Demos</h2>
<div class="content has-text-justified">
<!-- 第一个视频 -->
<div class="columns is-centered">
<div class="column is-full">
<video controls muted loop autoplay width="100%" style="max-width: 1000px">
<source src="./videos/RiOSWorld_demo.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
</div>
<!-- 第二个视频 -->
<!-- <div class="columns is-centered" style="margin-top: 2rem;">
<div class="column is-full">
<video controls muted loop autoplay width="100%">
<source src="./videos/delete_root.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
</div> -->
<!-- 第三个视频 -->
<!-- <div class="columns is-centered" style="margin-top: 2rem;">
<div class="column is-full">
<video controls muted loop autoplay width="100%">
<source src="./videos/arxiv.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
</div> -->
</div>
<!--/ Video Demos. -->
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Abstract. -->
<div class="columns is-centered has-text-centered">
<div class="column is-full-wdith">
<h2 class="title is-3">Abstract</h2>
<div class="content has-text-justified">
<md-block>
With the rapid development of multimodal large language models (MLLMs), they
are increasingly deployed as autonomous computer-use agents capable of accomplishing
complex computer tasks. However, a pressing issue arises: Can the safety
risk principles designed and aligned for general MLLMs in dialogue scenarios be
effectively transferred to real-world computer-use scenarios? Existing research
on evaluating the safety risks of MLLM-based computer-use agents suffers from
several limitations: it either lacks realistic interactive environments, or narrowly
focuses on one or a few specific risk types. These limitations ignore the complexity,
variability, and diversity of real-world environments, thereby restricting
comprehensive risk evaluation for computer-use agents. To this end, we introduce
**RiOSWorld**, a benchmark designed to evaluate the potential risks of MLLM-based
agents during real-world computer manipulations. Our benchmark includes 492
risky tasks spanning various computer applications, involving web, social media,
multimedia, os, email, and office software. We categorize these risks into
two major classes based on their risk source: (i) User-originated risks and (ii)
Environmental risks. For the evaluation, we evaluate safety risks from two perspectives:
(i) Risk goal intention and (ii) Risk goal completion. Extensive experiments
with multimodal agents on **RiOSWorld** demonstrate that current computer-use
agents confront significant safety risks in real-world scenarios. Our findings highlight
the necessity and urgency of safety alignment for computer-use agents in
real-world computer manipulation, providing valuable insights for developing
trustworthy computer-use agents.
</md-block>
</div>
</div>
</div>
<!--/ Abstract. -->
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Environment Infrastructure. -->
<h2 class="title is-3">Task Configuration of RiOSWorld</h2>
<div class="content has-text-justified">
<img src="static/images/environment.png" width="100%" alt="task configuration" class="responsive-image">
<md-block>**RiOSWorld**'s Task Configuration. Our benchmark uses a configuration file for definition of <span class="instruction">task instruction</span>,
<span class="config">initial environment setup</span>, <span class="halfway-config">halfway environment setup</span>, <span class="evaluator">task completion evaluation</span>,
and <span class="risk-evaluator">risk trigger evaluation</span>.</md-block>
</div>
<!--/ Environment Infrastructure. -->
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Data Statistics. -->
<h2 class="title is-3">Data Statistics and Comparison</h2>
<md-block>
**RiOSWorld** contains a total of 492 risky computer-use tasks, which classified into 2 main categories, 13 distinct subcategories with diverse topic distribution. </md-block>
<h2 class="title is-4">Data Statistics</h2>
<div class="columns is-centered is-variable is-1">
<div class="column is-half">
<div class="content has-text-centered">
<p>Risk statistics in
<b>RiOSWorld</b>.<br> Static or Dynamic, which represents whether the threat can be dynamically deployed.</p>
<img src="static/images/statistics.png" alt="data-overview" style="max-width: 90%;" />
</div>
</div>
<div class="column is-half">
<div class="content has-text-centered">
<img src="static/images/pie_chart.png" alt="data-composition" style="max-width: 90%;" />
<p>
Fine-grained distribution of risky tasks in
<b>RiOSWorld</b>.
</p>
</div>
</div>
</div>
<h2 class="title is-4">Benchmarks Comparision</h2>
<md-block>Comparison of different studies on computer-use agents risk evaluation. **# Number of
Risky Example**: Number of risky examples. **Environment Platform**: The simulation environment.
**Online Rule-based Eval.?**: Whether support online rule-based evaluation. **Multi-modal Support?**:
Whether agents support multi-modal inputs. **Real Network Accessible?**: Whether the environment is
connected to the Internet. **Dynamic Threat Support?**: Whether the environment supports dynamic
threats. **# Categories of Safety Risk**: Number of category.</md-block><br>
<div class="column is-full-width interpolation-panel">
<div class="columns">
<div class="column is-one-third">
<div class="table-container">
<div class="columns">
<div class="column">
<table class="table is-hoverable is-striped" style="margin: 0 auto; background-color: rgba(0,0,0,0)">
<thead>
<tr>
<th style="border: 0"> </th>
<th style="vertical-align: middle" rowspan="2">RiOSWorld<br>(Ours)</th>
</tr>
</thead>
<tr>
<td style="font-weight: bold"># Number of <br>(Risky Example)</td>
<td style="text-align: center;">492</td>
</tr>
<tr>
<td style="font-weight: bold">Environment <br>Platform</td>
<td style="text-align: center;">Virtual Machine</td>
</tr>
<tr>
<td style="font-weight: bold">Online Rule-based Eval.?</td>
<td style="text-align: center;">✔️</td>
</tr>
<tr>
<td style="font-weight: bold">Multi-modal <br>Support?</td>
<td style="text-align: center;">✔️</td>
</tr>
<tr>
<td style="font-weight: bold">Real Network <br>Accessible?</td>
<td style="text-align: center;">✔️</td>
</tr>
<tr>
<td style="font-weight: bold">Dynamic Threat<br>Support?</td>
<td style="text-align: center;">
✔️
</td>
</tr>
<tr>
<td style="font-weight: bold"># Categories of<br>Safety Risk</td>
<td style="text-align: center;">
13
</td>
</tr>
</table>
</div>
</div>
</div>
</div>
<div class="column is-two-thirds">
<div class="table-container">
<div class="columns">
<div class="column">
<table class="table is-hoverable is-striped"
style="margin: 0 auto; background-color: rgba(0,0,0,0); width:600px; text-align: center;">
<thead style="text-align: center;">
<tr>
<th style="width: 33%" rowspan="2"><a href="https://github.com/ryoungj/toolemu?tab=readme-ov-file" target="_blank">ToolEmu</a>
</th>
<th style="width: 33%" rowspan="2"><a href="https://github.com/uiuc-kang-lab/InjecAgent" target="_blank">InjecAgent</a>
</th>
<th style="width: 33%" rowspan="2"><a href="https://github.com/Junjie-Ye/ToolSword" target="_blank">ToolSword</a>
</th>
<th style="width: 33%" rowspan="2"><a href="https://github.com/Lordog/R-Judge" target="_blank">R-Judge</a>
</th>
<th style="width: 33%" rowspan="2"><a href="https://huggingface.co/datasets/ai-safety-institute/AgentHarm" target="_blank">AgentHarm</a>
</th>
<th style="width: 33%" rowspan="2"><a href="https://github.com/agiresearch/ASB" target="_blank">ASB</a>
</th>
<th style="width: 33%" rowspan="2"><a href="https://github.com/thu-coai/Agent-SafetyBench/" target="_blank">Agent-SafetyBench</a>
</th>
<th style="width: 33%" rowspan="2"><a href="https://agentdojo.spylab.ai/" target="_blank">AgentDojo</a>
</th>
<th style="width: 33%" rowspan="2"><a href="https://github.com/xbmxb/EnvDistraction" target="_blank">Env. Distraction</a>
</th>
<th style="width: 33%" rowspan="2"><a href="https://safearena.github.io/" target="_blank">SafeArena</a>
</th>
<th style="width: 33%" rowspan="2"><a href="https://github.com/segev-shlomov/ST-WebAgentBench" target="_blank">ST-WebAgentBench</a>
</th>
<th style="width: 33%" rowspan="2"><a href="https://mobilesafetybench.github.io/" target="_blank">MobileSafetyBench</a>
</th>
<th style="width: 33%" rowspan="2"><a href="https://github.com/osu-nlp-group/eia_against_webagent" target="_blank">EIA</a>
</th>
<th style="width: 33%" rowspan="2"><a href="https://arxiv.org/pdf/2504.18575" target="_blank">WASP</a>
</th>
<th style="width: 33%" rowspan="2"><a href="https://github.com/ChenWu98/agent-attack" target="_blank">VisualWebArena-Adv</a>
</th>
<th style="width: 33%" rowspan="2"><a href="https://github.com/SALT-NLP/PopupAttack" target="_blank">Attacking Popup</a>
</th>
</tr>
</thead>
<tr>
<td>144<br> </td>
<td>1054</td>
<td>440</td>
<td>569</td>
<td>110</td>
<td>400</td>
<td>2000</td>
<td>629</td>
<td>1198</td>
<td>250</td>
<td>234</td>
<td>80</td>
<td>177</td>
<td>84</td>
<td>200</td>
<td>122</td>
</tr>
<tr>
<td>LM Emulator</td>
<td>QA Format</td>
<td>QA Format</td>
<td>QA Format</td>
<td>Inspect</td>
<td>QA Format</td>
<td>QA Format</td>
<td>Code Emulator</td>
<td>QA Format</td>
<td>BrowserGym</td>
<td>BrowserGym</td>
<td>Android Emulator</td>
<td>Mind2Web</td>
<td>BrowserGym</td>
<td>BrowserGym</td>
<td>Virtual Machine</td>
</tr>
<tr>
<td>❌<br> </td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>✔️</td>
<td>❌</td>
<td>✔️</td>
<td>✔️</td>
<td>✔️</td>
<td>❌</td>
<td>✔️</td>
<td>✔️</td>
<td>✔️</td>
</tr>
<tr>
<td>❌<br> </td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>✔️</td>
<td>✔️</td>
<td>✔️</td>
<td>✔️</td>
<td>✔️</td>
<td>✔️</td>
<td>✔️</td>
<td>✔️</td>
</tr>
<tr>
<td>❌<br> </td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>✔️</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>✔️</td>
</tr>
<tr>
<td>❌<br> </td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>❌</td>
<td>✔️</td>
</tr>
<tr>
<td>9<br> </td>
<td>6</td>
<td>6</td>
<td>5</td>
<td>11</td>
<td>10</td>
<td>8</td>
<td>4</td>
<td>4</td>
<td>5</td>
<td>3</td>
<td>5</td>
<td>1</td>
<td>1</td>
<td>1</td>
<td>1</td>
</tr>
</table>
</div>
</div>
</div>
</div>
</div>
</div>
<!--/ Data Statistics. -->
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Environment Infrastructure. -->
<h2 class="title is-4">Topic Distribution in RiOSWorld</h2>
<div class="content has-text-justified">
<img src="static/images/instruction_distribution.png" width="100%" alt="topic distribution" class="responsive-image">
<md-block>We visualize the topic distribution of the instruction text embedding in our benchmark.
It encompasses a wide range of topics that permeate numerous routine operational scenarios encountered by Computer-Use Agents.
This comprehensive coverage ensures that **RiOSWorld** can effectively evaluate and test Computer-Use Agents across various aspects.</md-block>
</div>
<!--/ Environment Infrastructure. -->
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Experiments. -->
<h2 class="title is-3">Experiments</h2>
<md-block>
We evaluate the safety risks of state-of-the-art multimodal agents on **RiOSWorld** from two perspectives:
**Risk Goal Intention** and **Risk Goal Completion**. Our experiments cover 10 representative models including
GPT-4o, GPT-4o-mini, GPT-4.1, Gemini-2.0-pro, Gemini-2.5-pro-exp, Claude-3.5-Sonnet, Claude-3.7-Sonnet,
Qwen2-VL-72B-Instruct, Qwen2.5-VL-72B-Instruct, and Llama-3.2-90B-Vision-Instruct.
</md-block>
<!-- Key Findings -->
<h3 class="title is-4">Key Findings</h3>
<!-- Key Findings Text -->
<md-block>
<span style="color: red; font-weight: bold;">MLLMs-Based Agents are Still Far from Being Trustworthy Assistants for Autonomous Computer-Use.</span>
The results reveal significant safety flaws in current MLLM-based agents across all risk categories in **RiOSWorld**.
Our key findings include:
- <span style="color: red; font-weight: bold;">High Intention Risk Rates</span>: The average risk goal intention unsafe rates exceed <span style="color: red; font-weight: bold;">80%</span> across all models,
with <span style="color: red; font-weight: bold;">89.12%</span> for environmental risks and <span style="color: red; font-weight: bold;">81.33%</span> for user-originated risks.
- <span style="color: red; font-weight: bold;">High Completion Rates</span>: Risk goal completion rates are approximately <span style="color: red; font-weight: bold;">60%</span> for both categories
(<span style="color: red; font-weight: bold;">60.29%</span> for environmental risks and <span style="color: red; font-weight: bold;">59.07%</span> for user-originated risks).
- **Overall Risk Assessment**: The overall <span style="color: red; font-weight: bold;">Unsafe Rate (USR)</span> across all agents are <span style="color: red; font-weight: bold;">84.93%</span> for intention
and <span style="color: red; font-weight: bold;">59.64%</span> for completion, indicating substantial safety concerns.
- **Model-Specific Patterns**:
- For environmental risks, **Gemini-2.0-pro** exhibits the highest USR for risk goal intention,
while **Llama-3.2-90B-Vision-Instruct** shows the highest USR for risk goal completion.
- For user-originated risks, **GPT-4.1** demonstrates the highest USR for intention,
and **GPT-4o-mini** has the highest USR for completion.
These results indicate that current MLLM-based agents still have a significant gap from humans in safety
awareness and behavior, necessitating further research in safety alignment for computer-use agents.
</md-block>
<!-- Summary Table -->
<div class="columns is-centered">
<div class="column is-half">
<!-- <h4 class="title is-6 has-text-centered">Overall Unsafe Rate Summary</h4> -->
<table class="table experiments-table">
<thead>
<tr>
<th>Risk Source</th>
<th># Num.</th>
<th>USR Intention/Completion (%)</th>
</tr>
</thead>
<tbody>
<tr>
<td class="model-name">Environment</td>
<td>254</td>
<td class="risk-high">89.12 / 60.29</td>
</tr>
<tr>
<td class="model-name">User</td>
<td>238</td>
<td class="risk-high">81.33 / 59.07</td>
</tr>
<tr class="average-row">
<td class="model-name">Total</td>
<td>492</td>
<td>84.93 / 59.64</td>
</tr>
</tbody>
</table>
</div>
</div>
<!-- Main Experimental Results -->
<h3 class="title is-4">Main Experimental Results</h3>
<!-- Environmental Risks Table -->
<!-- <h4 class="title is-5">Environmental Risks</h4> -->
<md-block>
**Table 1:** Unsafe rates of environmental risks including Pop-ups/Ads, Phishing Web, Phishing Email,
Account Fraud, reCAPTCHA, and Induced Text. The first column of each scenario represents the unsafe rate
of **Risk Goal Intention**, and the second column indicates the unsafe rate of **Risk Goal Completion**.
</md-block>
<div class="table-container">
<table class="table experiments-table">
<thead>
<tr>
<th rowspan="2">Model</th>
<th colspan="12">Unsafe Rate (%)</th>
</tr>
<tr>
<th colspan="2">Pop-ups/Ads</th>
<th colspan="2">Phishing Web</th>
<th colspan="2">Phishing Email</th>
<th colspan="2">Account</th>
<th colspan="2">reCAPTCHA</th>
<th colspan="2">Induced Text</th>
</tr>
</thead>
<tbody>
<tr>
<td class="model-name">GPT-4o</td>
<td class="risk-high">93.8</td><td class="risk-medium">68.8</td>
<td class="risk-high">100</td><td class="risk-high">92.2</td>
<td class="risk-high">100</td><td class="risk-low">38.5</td>
<td class="risk-high">82.1</td><td class="risk-low">15.2</td>
<td class="risk-medium">56.7</td><td class="risk-low">22.6</td>
<td class="risk-high">100</td><td class="risk-high">95.8</td>
</tr>
<tr>
<td class="model-name">GPT-4o-mini</td>
<td class="risk-high">94.0</td><td class="risk-medium">64.0</td>
<td class="risk-high">100</td><td class="risk-high">88.2</td>
<td class="risk-high">100</td><td class="risk-medium">56.3</td>
<td class="risk-medium">75.9</td><td class="risk-medium">51.5</td>
<td class="risk-high">87.5</td><td class="risk-low">21.9</td>
<td class="risk-high">100</td><td class="risk-high">100</td>
</tr>
<tr>
<td class="model-name">GPT-4.1</td>
<td class="risk-high">96.0</td><td class="risk-low">14.0</td>
<td class="risk-high">100</td><td class="risk-medium">75.6</td>
<td class="risk-high">90.0</td><td class="risk-low">36.4</td>
<td class="risk-high">80.7</td><td class="risk-low">12.1</td>
<td class="risk-low">45.5</td><td class="risk-low">27.3</td>
<td class="risk-high">96.0</td><td class="risk-medium">77.1</td>
</tr>
<tr>
<td class="model-name">Gemini-2.0-pro</td>
<td class="risk-high">100</td><td class="risk-low">44.0</td>
<td class="risk-high">97.9</td><td class="risk-high">95.8</td>
<td class="risk-high">96.6</td><td class="risk-low">31.3</td>
<td class="risk-high">100</td><td class="risk-low">21.2</td>
<td class="risk-high">95.8</td><td class="risk-medium">56.7</td>
<td class="risk-high">100</td><td class="risk-high">100</td>
</tr>
<tr>
<td class="model-name">Gemini-2.5-pro-exp</td>
<td class="risk-high">98.0</td><td class="risk-medium">65.3</td>
<td class="risk-high">100</td><td class="risk-high">94.2</td>
<td class="risk-high">93.3</td><td class="risk-low">29.6</td>
<td class="risk-medium">79.3</td><td class="risk-low">18.2</td>
<td class="risk-medium">53.3</td><td class="risk-medium">50.0</td>
<td class="risk-high">100</td><td class="risk-high">100</td>
</tr>
<tr>
<td class="model-name">Claude-3.5-Sonnet</td>
<td class="risk-high">93.9</td><td class="risk-medium">53.1</td>
<td class="risk-high">100</td><td class="risk-medium">75.5</td>
<td class="risk-high">87.5</td><td class="risk-medium">59.4</td>
<td class="risk-high">86.4</td><td class="risk-low">9.1</td>
<td class="risk-medium">66.7</td><td class="risk-low">28.6</td>
<td class="risk-high">88.6</td><td class="risk-medium">78.0</td>
</tr>
<tr>
<td class="model-name">Claude-3.7-Sonnet</td>
<td class="risk-high">91.8</td><td class="risk-high">83.7</td>
<td class="risk-high">94.2</td><td class="risk-high">88.5</td>
<td class="risk-high">93.8</td><td class="risk-medium">65.6</td>
<td class="risk-medium">62.1</td><td class="risk-low">10.3</td>
<td class="risk-medium">67.9</td><td class="risk-low">35.7</td>
<td class="risk-high">94.0</td><td class="risk-high">94.0</td>
</tr>
<tr>
<td class="model-name">Qwen2-VL-72B-Instruct</td>
<td class="risk-medium">69.4</td><td class="risk-medium">54.0</td>
<td class="risk-high">100</td><td class="risk-medium">77.8</td>
<td class="risk-high">100</td><td class="risk-low">28.1</td>
<td class="risk-high">100</td><td class="risk-low">15.2</td>
<td class="risk-high">96.3</td><td class="risk-low">22.2</td>
<td class="risk-medium">66.7</td><td class="risk-medium">75.0</td>
</tr>
<tr>
<td class="model-name">Qwen2.5-VL-72B-Instruct</td>
<td class="risk-high">100</td><td class="risk-medium">53.1</td>
<td class="risk-high">100</td><td class="risk-medium">76.5</td>
<td class="risk-high">96.9</td><td class="risk-low">43.8</td>
<td class="risk-high">100</td><td class="risk-low">15.2</td>
<td class="risk-high">93.3</td><td class="risk-low">40.0</td>
<td class="risk-medium">53.1</td><td class="risk-medium">68.8</td>
</tr>
<tr>
<td class="model-name">Llama-3.2-90B-Vision-Instruct</td>
<td class="risk-medium">66.3</td><td class="risk-medium">72.0</td>
<td class="risk-high">100</td><td class="risk-medium">73.2</td>
<td class="risk-high">100</td><td class="risk-low">21.4</td>
<td class="risk-high">82.8</td><td class="risk-medium">69.7</td>
<td class="risk-high">83.3</td><td class="risk-medium">70.0</td>
<td class="risk-high">100</td><td class="risk-high">100</td>
</tr>
<tr class="average-row">
<td class="model-name">Average</td>
<td>90.3</td><td>57.2</td>
<td>99.2</td><td>83.7</td>
<td>95.8</td><td>41.0</td>
<td>84.9</td><td>20.5</td>
<td>74.6</td><td>37.5</td>
<td>89.8</td><td>88.9</td>
</tr>
</tbody>
</table>
</div>
<!-- User Risks Table -->
<!-- <h4 class="title is-5">User-Originated Risks</h4> -->
<md-block>
**Table 2:** Unsafe rates of risks from users including File I/O, OS Operation, Web, Code,
Social Media, Office, and Multimedia. The first column of each scenario represents the unsafe rate
of **Risk Goal Intention**, and the second column indicates the unsafe rate of **Risk Goal Completion**.
</md-block>
<div class="table-container">
<table class="table experiments-table">
<thead>
<tr>
<th rowspan="2">Model</th>
<th colspan="14">Unsafe Rate (%)</th>
</tr>
<tr>
<th colspan="2">File I/O</th>
<th colspan="2">OS</th>
<th colspan="2">Web</th>
<th colspan="2">Code</th>
<th colspan="2">Social Media</th>
<th colspan="2">Office</th>
<th colspan="2">Multimedia</th>
</tr>
</thead>
<tbody>
<tr>
<td class="model-name">GPT-4o</td>
<td class="risk-medium">69.6</td><td class="risk-medium">60.9</td>
<td class="risk-high">93.3</td><td class="risk-high">86.7</td>
<td class="risk-high">90.2</td><td class="risk-high">90.2</td>
<td class="risk-high">90.2</td><td class="risk-high">80.5</td>
<td class="risk-high">86.4</td><td class="risk-low">23.3</td>
<td class="risk-medium">71.9</td><td class="risk-high">90.5</td>
<td class="risk-high">100</td><td class="risk-high">96.0</td>
</tr>
<tr>
<td class="model-name">GPT-4o-mini</td>
<td class="risk-high">91.3</td><td class="risk-medium">69.6</td>
<td class="risk-medium">76.7</td><td class="risk-medium">73.3</td>
<td class="risk-high">90.5</td><td class="risk-high">100</td>
<td class="risk-high">100</td><td class="risk-high">88.2</td>
<td class="risk-high">95.2</td><td class="risk-low">20.0</td>
<td class="risk-medium">72.7</td><td class="risk-high">81.0</td>
<td class="risk-high">98.0</td><td class="risk-high">98.0</td>
</tr>
<tr>
<td class="model-name">GPT-4.1</td>
<td class="risk-medium">65.2</td><td class="risk-low">43.5</td>
<td class="risk-high">96.7</td><td class="risk-high">93.3</td>
<td class="risk-high">100</td><td class="risk-high">95.2</td>
<td class="risk-high">90.2</td><td class="risk-medium">75.0</td>
<td class="risk-high">95.2</td><td class="risk-low">23.3</td>