-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
771 lines (740 loc) Β· 37 KB
/
index.html
File metadata and controls
771 lines (740 loc) Β· 37 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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<meta name="description" content="π-Teaming: Multi-Turn Jailbreaks and Defenses with Adaptive Multi-Agents">
<meta name="keywords" content="π-Teaming, Language Models, Safety, Multi-turn Jailbreaks">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>π-Teaming: Multi-Turn Jailbreaks and Defenses with Adaptive Multi-Agents</title>
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Noto+Sans+Math|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/storm.png">
<script src="https://unpkg.com/marked@15.0.8/marked.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>
</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">
<span class="x-title">π</span>-Teaming: Multi-Turn Jailbreaks and Defenses with Adaptive Multi-Agents
</h1>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://salmanrahman.net/" target="_blank" rel="noopener noreferrer">Salman Rahman</a><sup>1*</sup>,</span>
<span class="author-block">
<a href="https://liweijiang.me/" target="_blank" rel="noopener noreferrer">Liwei Jiang</a><sup>2*</sup>,</span>
<span class="author-block">
<a href="https://www.jshiffer.xyz/" target="_blank" rel="noopener noreferrer">James Shiffer</a><sup>1*</sup>,</span>
<span class="author-block">
<a href="https://genglinliu.github.io/" target="_blank" rel="noopener noreferrer">Genglin Liu</a><sup>1</sup>,</span>
<span class="author-block">
<a href="https://www.linkedin.com/in/sheriff-issaka" target="_blank" rel="noopener noreferrer">Sheriff Issaka</a><sup>1</sup>,</span>
<span class="author-block">
<a href="https://rizwan09.github.io/" target="_blank" rel="noopener noreferrer">Md Rizwan Parvez</a><sup>3</sup>,</span>
<span class="author-block">
<a href="https://www.hamidpalangi.com/" target="_blank" rel="noopener noreferrer">Hamid Palangi</a><sup>4</sup>,</span>
<span class="author-block">
<a href="https://web.cs.ucla.edu/~kwchang/" target="_blank" rel="noopener noreferrer">Kai-Wei Chang</a><sup>1</sup>,</span>
<span class="author-block">
<a href="https://yejinc.github.io/" target="_blank" rel="noopener noreferrer">Yejin Choi</a><sup>5</sup>,</span>
<span class="author-block">
<a href="https://saadiagabriel.com/" target="_blank" rel="noopener noreferrer">Saadia Gabriel</a><sup>1</sup></span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>1</sup>University of California, Los Angeles,</span>
<span class="author-block"><sup>2</sup>University of Washington,</span>
<span class="author-block"><sup>3</sup>Qatar Computing Research Institute,</span>
<span class="author-block"><sup>4</sup>Google,</span>
<span class="author-block"><sup>5</sup>Stanford University</span>
</div>
<div class="is-size-6 publication-authors"><sup>*</sup>Equal contribution</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<span class="link-block">
<a href="https://arxiv.org/pdf/2504.13203" target="_blank" rel="noopener noreferrer"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>Paper</span>
</a>
</span>
<span class="link-block">
<a href="https://arxiv.org/abs/2504.13203" target="_blank" rel="noopener noreferrer"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
</span>
<!-- Code Link. -->
<span class="link-block">
<a href="https://github.com/salman-lui/x-teaming" target="_blank" rel="noopener noreferrer"
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>
<!-- Dataset Link. -->
<span class="link-block">
<a href="https://huggingface.co/datasets/marslabucla/XGuard-Train" target="_blank" rel="noopener noreferrer"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-database"></i>
</span>
<span>Dataset</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<div class="content has-text-justified">
<p>
Multi-turn interactions with language models (LMs) pose critical safety risks, as harmful intent can be strategically spread across exchanges.
Yet, the vast majority of prior work has focused on single-turn safety, while adaptability and diversity remain among the key challenges of multi-turn red-teaming.
To address these challenges, we present <span class="has-text-weight-bold"><span class="x-struck">π</span>-Teaming</span>, a scalable framework that systematically explores how seemingly harmless interactions escalate into harmful outcomes and generates corresponding attack scenarios.
</p>
<p>
<span class="x-struck">π</span>-Teaming employs collaborative agents for planning, attack optimization, and verification, achieving state-of-the-art multi-turn jailbreak effectiveness and diversity with <span class="has-text-weight-bold">success rates up to 99.4%</span> across representative leading open-weight and closed-source models. In particular, <span class="x-struck">π</span>-Teaming achieves a 96.2% attack success rate against the latest Claude 3.7 Sonnet model, which has been considered nearly immune to single-turn attacks.
</p>
<p>
Building on <span class="x-struck">π</span>-Teaming, we introduce <span class="x-guard has-text-weight-bold"><span class="x-struck">π</span>Guard-Train</span>, an open-source multi-turn safety training dataset that's ~20Γ larger than the previous best resource, comprising 30K interactive jailbreaks, designed to enable robust multi-turn safety alignment for LMs. Our work offers essential tools and insights for mitigating sophisticated conversational attacks, advancing the multi-turn safety of LMs.
</p>
</div>
</div>
</section>
<section class="section">
<div class="container is-max-desktop">
<!-- Framework Figure. -->
<div class="columns is-centered has-text-centered">
<div class="column is-full-width">
<img src="./static/images/framework.png" alt="π-Teaming Framework"
class="framework-image"/>
</div>
</div>
<!--/ Framework Figure. -->
</div>
</section>
<!-- Results Table Section -->
<section class="section" id="results">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-full-width">
<h2 id="results" class="title is-3 has-text-centered">Results <a class="is-size-5" href="#results"><i class="fas fa-link"></i></a></h2>
<div class="content has-text-centered">
<p class="table-caption">Attack success rate (ASR; %) on HarmBench test set.</p>
<div class="table-container">
<table class="table">
<thead>
<tr>
<th class="main-header header-dark"></th>
<th colspan="3" class="text-center main-header">Closed-Source</th>
<th colspan="5" class="text-center main-header header-dark">Open-Weight</th>
</tr>
<tr>
<th>Method</th>
<th class="small-header">GPT-4o</th>
<th class="small-header">Claude 3.5 Sonnet</th>
<th class="small-header">Gemini 2.0 Flash</th>
<th class="small-header">Llama 3-8B-IT</th>
<th class="small-header">Llama 3-70B-IT</th>
<th class="small-header">Llama 3-8B-IT (SafeMTData)</th>
<th class="small-header">Deepseek V3</th>
<th class="small-header">Qwen 2.5-32B-IT</th>
</tr>
</thead>
<tbody>
<tr class="category-row">
<td colspan="9"><em>Single-turn Methods</em></td>
</tr>
<tr>
<td><a href="https://arxiv.org/pdf/2307.15043">GCG</a></td>
<td>12.5</td>
<td>3.0</td>
<td>--</td>
<td>34.5</td>
<td>17.0</td>
<td>--</td>
<td>--</td>
<td>--</td>
</tr>
<tr>
<td><a href="https://arxiv.org/pdf/2310.08419">PAIR</a></td>
<td>39.0</td>
<td>3.0</td>
<td>--</td>
<td>18.7</td>
<td>36.0</td>
<td>--</td>
<td>--</td>
<td>--</td>
</tr>
<tr>
<td><a href="https://arxiv.org/pdf/2206.00052">CodeAttack</a></td>
<td>70.5</td>
<td>39.5</td>
<td>--</td>
<td>46.0</td>
<td>66.0</td>
<td>--</td>
<td>--</td>
<td>--</td>
</tr>
<tr class="category-row">
<td colspan="9"><em>Multi-turn Methods</em></td>
</tr>
<tr>
<td><a href="https://arxiv.org/pdf/2502.11054">RACE</a></td>
<td>82.8</td>
<td>--</td>
<td>--</td>
<td>--</td>
<td>--</td>
<td>--</td>
<td>--</td>
<td>--</td>
</tr>
<tr>
<td><a href="https://arxiv.org/pdf/2405.05610">CoA</a></td>
<td>17.5</td>
<td>3.4</td>
<td>--</td>
<td>25.5</td>
<td>18.8</td>
<td>--</td>
<td>--</td>
<td>--</td>
</tr>
<tr>
<td><a href="https://crescendo-the-multiturn-jailbreak.github.io/assets/pdf/CrescendoFullPaper.pdf">Crescendo</a></td>
<td>46.0</td>
<td>50.0</td>
<td>--</td>
<td>60.0</td>
<td>62.0</td>
<td>12.0</td>
<td>--</td>
<td>--</td>
</tr>
<tr>
<td><a href="https://arxiv.org/pdf/2410.10700">ActorAttack</a></td>
<td>84.5</td>
<td>66.5</td>
<td>42.1</td>
<td>79.0</td>
<td><strong>85.5</strong></td>
<td>21.4</td>
<td>68.6</td>
<td>--</td>
</tr>
<tr class="highlight-row">
<td><strong><span class="x-struck">π</span>-Teaming (Ours)</strong></td>
<td><strong>94.3</strong></td>
<td><strong>67.9*</strong></td>
<td><strong>87.4</strong></td>
<td><strong>85.5</strong></td>
<td>84.9</td>
<td><strong>91.8</strong></td>
<td><strong>98.1</strong></td>
<td><strong>99.4</strong></td>
</tr>
</tbody>
</table>
</div>
<p class="table-note">
* With full configuration (50 plans, 5 TextGrad tries, 10 turns), Claude 3.5 Sonnet achieves 67.9% and Claude 3.7 Sonnet achieves 96.2% ASR.
All other models, i.e. GPT-4o, Llama 3-IT variants, Gemini 2.0-Flash, Deepseek V3, and Qwen2.5-IT achieve 100% ASR on the HarmBench validation set. IT = Instruct.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- Example Conversation Section -->
<section class="section" id="examples">
<div class="container is-fullhd">
<h2 id="examples" class="title is-3 has-text-centered">Example Attacks <a class="is-size-5" href="#examples"><i class="fas fa-link"></i></a></h2>
<div class="container is-max-desktop">
<div class="notification has-background-danger-light has-text-centered">
<h4 class="is-4 has-text-danger"><strong>β οΈ Warning β οΈ</strong> The following dialogues contain highly offensive and disturbing content.</h4>
</div>
<br>
</div>
<div class="columns conversation-cols">
<!-- Left Panel -->
<div class="column is-one-fifth model-list">
<h2 class="title is-4 has-text-centered mb-4">Models</h2>
<div id="model-list">
<!-- Model items will be populated by JavaScript -->
</div>
</div>
<!-- Right Panel -->
<div class="column is-four-fifths conversation-panel">
<div class="tabs is-centered is-flex-shrink-0">
<ul>
<li class="is-active" data-mode="no-jailbreak"><a>Single Turn</a></li>
<li data-mode="x-teaming"><a><span class="x-struck">π</span>-Teaming</a></li>
</ul>
</div>
<div id="conversation-container">
<!-- Conversation messages will be populated dynamically -->
</div>
</div>
</div>
</div>
</section>
<section class="section" id="insights">
<div class="container is-max-desktop">
<div class="content has-text-justified">
<h2 id="insights" class="title is-3 has-text-centered">Key Insights <a class="is-size-5" href="#insights"><i class="fas fa-link"></i></a></h2>
</div>
</div>
<!-- Diversity -->
<div class="container is-max-desktop">
<div class="content has-text-justified">
<h2 class="subtitle is-4 has-text-centered">1. Plan and Attack Diversity</h2>
</div>
<div class="content has-text-justified">
<div class="has-text-centered my-8">
<img src="static/images/final_diversity_comparison_v2.png" alt="Diversity comparison between ActorAttack and π-Teaming">
</div>
<p>
<span class="x-struck">π</span>-Teaming also demonstrates significant improvements in attack diversity compared with
semantic-driven (<a href="https://arxiv.org/pdf/2411.15720">Chain of Attack</a>, <a href="https://arxiv.org/pdf/2410.10700">ActorAttack</a>)
and template-based (<a href="https://crescendo-the-multiturn-jailbreak.github.io/">Crescendo</a>) multi-turn jailbreak methods,
emulating the strategic diversity of human red-teamers. Compared with ActorAttack, the previous strongest multi-turn attack,
<span class="x-struck">π</span>-Teaming achieves a <span class="has-text-weight-bold">153% improvement</span> in attack plan
diversity and a <span class="has-text-weight-bold">62% improvement</span> in attack execution diversity, as measured by pairwise embedding similarities.
</p>
<div class="has-text-centered my-8">
<img src="static/images/diversity_plan.png" alt="Diversity of attack strategies">
</div>
<p>
The heatmap indicates that <span class="x-struck">π</span>-Teaming's method results in high mean diversity scores (average shown here = 0.82) between several plans for the same behavior,
where diversity scores are calculated as <code>1 - (cosine similarity of embeddings)</code>.
By creating strategically diverse plans with varied personas, distinct contexts, and approaches, we enable exploration of multiple attack scenarios for the same harmful behavior.
(For this plot, the 10 most dissimilar plans out of 50 were chosen.)
</p>
</div>
</div>
</section>
<!-- Ablation Testing -->
<section class="section">
<div class="container is-max-desktop">
<div class="content has-text-justified">
<h2 class="subtitle is-4 has-text-centered">2. Ablation Study</h2>
</div>
<div class="content has-text-justified">
<div class="has-text-centered my-8">
<img src="static/images/ablation_plot_three_compact.png" alt="Ablation tests">
</div>
<p>
For the ablation tests, we varied one parameter of our method (the maximum number of attack plans, conversation turns, or attempts of TextGrad optimization allowed) while keeping the others fixed.
The default number of attack plans was 10, and the default number of conversation turns was 7. Unless stated otherwise, TextGrad was disabled (0 attempts).
</p>
<p>
The main takeaways are that:
<ol>
<li>Based on the first chart, optimal performance requires sufficient strategy diversity, but additional plans beyond a certain point do not yield further improvements.</li>
<li>While multi-turn attacks are essential for overcoming safety defenses, longer conversations may cause context dilution as both attacker and target model manage an increasingly long message history.</li>
<li>Prompt optimization with <a href="https://textgrad.com/">TextGrad</a> substantially enhances attack success, while also validating our execution logic that stops optimization once the verification score improves over the previous turn, making the 3rd and 4th optimization attempts often unnecessary.</li>
</ol>
</p>
<p>
In practice, we found that increasing TextGrad attempts (3 β 4), conversation turns (6 β 7), and attack plans (5 β 10)
significantly improves attack success rates against the HarmBench validation set across models.
Thus, we adopted a configuration of 4 TextGrad attempts, 7 turns, and 10 plans for our main experiments.
</p>
</div>
</div>
</section>
<!-- Efficiency -->
<section class="section">
<div class="container is-max-desktop">
<div class="content has-text-justified">
<h2 class="subtitle is-4 has-text-centered">3. Efficiency of Successful Attacks</h2>
</div>
<div class="table-container has-text-centered">
<table class="table">
<thead>
<tr>
<th class="main-header header-dark"></th>
<th colspan="4" class="main-header">Closed-Source</th>
<th colspan="4" class="main-header header-dark">Open-Weight</th>
</tr>
<tr>
<th></th>
<th class="small-header">GPT-4o</th>
<th class="small-header">Claude 3.5 Sonnet</th>
<th class="small-header">Gemini 2.0 Flash</th>
<th class="small-header">Llama 3-8B-IT</th>
<th class="small-header">Llama 3-70B-IT</th>
<th class="small-header">Deepseek V3</th>
</tr>
</thead>
<tbody>
<tr>
<td>Avg. Turns</td>
<td>4.31</td>
<td>3.39</td>
<td>3.96</td>
<td>4.55</td>
<td>4.52</td>
<td>4.00</td>
</tr>
<tr>
<td>Avg. Plans</td>
<td>1.61</td>
<td>11.0</td>
<td>2.20</td>
<td>2.71</td>
<td>2.14</td>
<td>1.34</td>
</tr>
<tr>
<td>Avg. TextGrad</td>
<td>0.38</td>
<td>0.24</td>
<td>0.22</td>
<td>1.40</td>
<td>1.20</td>
<td>0.30</td>
</tr>
<tr>
<td>Avg. Attacker Tokens<br>(Model Limit)</td>
<td>1,470<br>(128K)</td>
<td>3,328<br>(200K)</td>
<td>1,884<br>(1M)</td>
<td>1,746<br>(8K)</td>
<td>1,311<br>(8K)</td>
<td>1,237<br>(128K)</td>
</tr>
</tbody>
</table>
</div>
<div class="content has-text-justified">
<p>
Successful attacks required approximately 4 conversation turns across models, with Claude requiring the most resources and open-weight models like Deepseek V3 requiring the fewest plans.
All attacks used only a small fraction of available context windows, demonstrating that <span class="x-struck">π</span>-Teaming effectively balances attack success with resource efficiency.
</p>
</div>
</div>
</section>
<!-- Agreement analysis -->
<section class="section">
<div class="container is-max-desktop">
<div class="content has-text-justified">
<h2 class="subtitle is-4 has-text-centered">4. Verifier Agreement</h2>
</div>
<div class="content has-text-justified">
<div class="has-text-centered my-8">
<img src="static/images/model_agreement_comparison_final.png" alt="Agreement rates between GPT-4o Verifier and others">
</div>
<p>
Because our experiments all used GPT-4o to verify successful jailbreak attempts, we conducted an agreement analysis across multiple evaluation methods.
GPT-4o has been used in prior multi-turn attack research<a href="https://arxiv.org/pdf/2410.10700"><sup>[1]</sup></a><a href="https://arxiv.org/pdf/2502.11054"><sup>[2]</sup></a>,
but it is possible that <a href="https://arxiv.org/pdf/2404.13076">LLM-based verifiers may bias results</a>.
</p>
<p>
Our analysis reveals strong overall agreement with <a href="https://arxiv.org/pdf/2402.04249">HarmBench</a> test classifiers (84.5% average), which themselves demonstrate 93.2% agreement with human evaluation.
<a href="https://arxiv.org/pdf/2312.06674">LlamaGuard 3</a> shows slightly lower agreement (69.09% average), consistent with previous findings on HarmBench test sets.
These substantial agreement rates with HarmBench test classifiers support our use of GPT-4o as a verifier for this benchmark.
</p>
</div>
</div>
</section>
<!-- XGuard-Train -->
<section class="section" id="xguard-train">
<div class="container is-max-desktop">
<div class="content has-text-justified">
<h2 class="title is-3 has-text-centered" id="xguard-train"><span class="x-title">π</span>Guard-Train <a class="is-size-5" href="#xguard-train"><i class="fas fa-link"></i></a></h2>
<div class="has-text-centered">
<a href="https://huggingface.co/datasets/marslabucla/XGuard-Train">
<img alt="Dataset on HF" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/dataset-on-hf-md.svg">
</a>
</div>
</div>
<div class="content has-text-justified">
<p>
This dataset consists of 30,695 multi-turn conversations, with complete attack-refusal pairs that enable robust multi-turn safety training.
</p>
<p>
We constructed <span class="x-struck">π</span>Guard-Train by proportionately sampling 10,000 harmful behaviors from <a href="https://huggingface.co/datasets/allenai/wildjailbreak">WildJailbreak's</a> vanilla harmful category.
For each harmful behavior, our planner generated between two to five distinct attack plans, resulting in diverse attack trajectories incorporating various personas, contexts, and conversation approaches.
We executed these plans using the complete <span class="x-struck">π</span>-Teaming pipeline, with GPT-4o, Gemini 2.0 Flash, and Deepseek V3 as target models, and Qwen-2.5-32B-IT handling both attack execution and TextGrad optimization.
The pipeline refined attacker queries when verification scores decreased and dynamically adjusted plans that failed to achieve their harmful targets. This process resulted in highly effective jailbreaking conversations with an average of 5.10 turns, where one turn represents an attacker prompt and target model response pair.
For successful jailbreaks, we replaced harmful model responses with carefully crafted helpful refusals.
</p>
</div>
</div>
</section>
<!-- Safety Model Results -->
<section class="section" id="safety-results">
<div class="container is-max-desktop">
<div class="content has-text-justified">
<h3 class="subtitle is-4 has-text-centered">Safety Model Results</h3>
<p class="has-text-centered">
Evaluation of models fine-tuned with <span class="x-struck">π</span>Guard-Train, <a href="https://huggingface.co/datasets/allenai/tulu-v2-sft-mixture">TuluMix</a>,
and <a href="https://huggingface.co/datasets/SafeMTData/SafeMTData">SafeMT</a> across multi-turn safety, single-turn safety, and general capabilities.
</p>
<div class="table-container">
<style>
.section-border-right {
border-right: 2px solid #dbdbdb !important;
}
.header-section {
background-color: #f5f5f5 !important;
font-weight: bold;
}
.metric-header {
border-bottom: 2px solid #dbdbdb !important;
}
.table td, .table th {
vertical-align: middle !important;
}
</style>
<table class="table is-bordered is-narrow is-hoverable is-fullwidth">
<thead>
<tr>
<th rowspan="2" class="has-text-centered header-section">Model</th>
<th colspan="3" class="has-text-centered header-section section-border-right">Multi-Turn Safety (ASR β)</th>
<th colspan="3" class="has-text-centered header-section section-border-right">Single-Turn Safety (ASR β)</th>
<th colspan="4" class="has-text-centered header-section">Capability (Accuracy β)</th>
</tr>
<tr>
<th class="has-text-centered"><span class="x-struck">π</span>-Team<br>(Ours)</th>
<th class="has-text-centered">Actor<br>Attack</th>
<th class="has-text-centered section-border-right">Average</th>
<th class="has-text-centered">DAN<sup>a</sup></th>
<th class="has-text-centered">WildGuard<sup>b</sup><br>Adv/Van</th>
<th class="has-text-centered section-border-right">XS<br>Test<sup>c</sup></th>
<th class="has-text-centered">MMLU</th>
<th class="has-text-centered">GSM8K</th>
<th class="has-text-centered">MATH</th>
<th class="has-text-centered">GPQA</th>
</tr>
</thead>
<tbody>
<tr>
<td colspan="11" class="model-row"><em>Llama-3.1-8B</em></td>
</tr>
<tr>
<td>TuluMix</td>
<td class="has-text-centered">80.5</td>
<td class="has-text-centered">44.0</td>
<td class="has-text-centered section-border-right">62.3</td>
<td class="has-text-centered"><strong>2.3</strong></td>
<td class="has-text-centered">25.8/<strong>6.7</strong></td>
<td class="has-text-centered section-border-right"><strong>24.0</strong></td>
<td class="has-text-centered">0.65</td>
<td class="has-text-centered">0.59</td>
<td class="has-text-centered">0.14</td>
<td class="has-text-centered">0.24</td>
</tr>
<tr>
<td>TuluMix + SafeMT</td>
<td class="has-text-centered">93.7<sup>*</sup></td>
<td class="has-text-centered"><strong>8.9</strong></td>
<td class="has-text-centered section-border-right">51.3</td>
<td class="has-text-centered">11.3</td>
<td class="has-text-centered">27.3/7.3</td>
<td class="has-text-centered section-border-right">28.7</td>
<td class="has-text-centered">0.65</td>
<td class="has-text-centered">0.57</td>
<td class="has-text-centered">0.14</td>
<td class="has-text-centered">0.26</td>
</tr>
<tr class="highlight-row">
<td class="has-text-weight-bold">TuluMix + <span class="x-struck">π</span>Guard</td>
<td class="has-text-centered"><strong>52.2<sup>*</sup></strong></td>
<td class="has-text-centered">18.9</td>
<td class="has-text-centered section-border-right"><strong>35.6</strong></td>
<td class="has-text-centered">8.3</td>
<td class="has-text-centered"><strong>23.7</strong>/7.5</td>
<td class="has-text-centered section-border-right">28.0</td>
<td class="has-text-centered"><strong>0.65</strong></td>
<td class="has-text-centered"><strong>0.59</strong></td>
<td class="has-text-centered"><strong>0.14</strong></td>
<td class="has-text-centered"><strong>0.28</strong></td>
</tr>
<tr>
<td colspan="11" class="model-row"><em>Qwen-2.5-7B</em></td>
</tr>
<tr>
<td>TuluMix</td>
<td class="has-text-centered">79.2</td>
<td class="has-text-centered">21.4</td>
<td class="has-text-centered section-border-right">50.3</td>
<td class="has-text-centered"><strong>1.0</strong></td>
<td class="has-text-centered">27.3/<strong>10.0</strong></td>
<td class="has-text-centered section-border-right">34.9</td>
<td class="has-text-centered"><strong>0.74</strong></td>
<td class="has-text-centered"><strong>0.70</strong></td>
<td class="has-text-centered">0.15</td>
<td class="has-text-centered">0.31</td>
</tr>
<tr>
<td>TuluMix + SafeMT</td>
<td class="has-text-centered">77.4</td>
<td class="has-text-centered"><strong>8.8</strong></td>
<td class="has-text-centered section-border-right">43.1</td>
<td class="has-text-centered">4.3</td>
<td class="has-text-centered"><strong>26.1</strong>/11.2</td>
<td class="has-text-centered section-border-right">36.2</td>
<td class="has-text-centered">0.73</td>
<td class="has-text-centered">0.33</td>
<td class="has-text-centered"><strong>0.19</strong></td>
<td class="has-text-centered">0.32</td>
</tr>
<tr class="highlight-row">
<td class="has-text-weight-bold">TuluMix + <span class="x-struck">π</span>Guard</td>
<td class="has-text-centered"><strong>40.9</strong></td>
<td class="has-text-centered">18.2</td>
<td class="has-text-centered section-border-right"><strong>29.6</strong></td>
<td class="has-text-centered">1.6</td>
<td class="has-text-centered">28.8/13.1</td>
<td class="has-text-centered section-border-right"><strong>27.8</strong></td>
<td class="has-text-centered"><strong>0.74</strong></td>
<td class="has-text-centered">0.63</td>
<td class="has-text-centered">0.16</td>
<td class="has-text-centered"><strong>0.33</strong></td>
</tr>
</tbody>
</table>
</div>
<div class="content is-size-7 has-text-centered">
<p>
<span class="x-struck">π</span>Guard models were trained on a 14K subset of the full dataset.<br>
<sup>a</sup>DAN: do anything now;
<sup>b</sup>WildGuard: Adv = Adversarial Harm, Van = Vanilla Harm;
<sup>c</sup>XS Test shows refusal accuracy values converted to (100 - original score).<br>
<sup>*</sup>Results use full configuration (50 plans, 5 TextGrad tries, 10 turns).
</p>
</div>
<div class="has-text-justified">
<p>
Our <span class="x-struck">π</span>Guard-Train-tuned model maintains the <span class="has-text-weight-bold">best average performance</span> across both multi-turn attack methodologies.
For single-turn safety benchmarks, the <span class="x-struck">π</span>Guard-Train-tuned model performs well in protecting against
adversarial harm in the WildGuard benchmark (23.7%), outperforming both SafeMTData (27.3%)
and baseline (25.8%) models, while also maintaining low ASR in other single-turn benchmarks like
Do Anything Now (DAN) and XSTest. Our <span class="x-struck">π</span>Guard-Train-tuned model <span class="has-text-weight-bold">preserves general capabilities</span>
across all benchmarks (MMLU, GSM8K, MATH, and GPQA), with GPQA showing improvement (0.28 vs. 0.26 for SafeMTData and 0.24 for TuluMix).
Similar trends appear in our evaluations with Qwen-2.5-7B.
</p>
</div>
</div>
</div>
</section>
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>@inproceedings{
rahman2025xteaming,
title={X-Teaming: Multi-Turn Jailbreaks and Defenses with Adaptive Multi-Agents},
author={Salman Rahman and Liwei Jiang and James Shiffer and Genglin Liu and Sheriff Issaka and Md Rizwan Parvez and Hamid Palangi and Kai-Wei Chang and Yejin Choi and Saadia Gabriel},
booktitle={Second Conference on Language Modeling},
year={2025},
url={https://openreview.net/forum?id=gKfj7Jb1kj}
}</code></pre>
</div>
</section>
<!-- Organization Logos Section -->
<section class="section" id="org-logos">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<div class="org-logos" style="margin: 2rem 0;">
<a href="https://www.ucla.edu/" target="_blank" rel="noopener noreferrer">
<img src="static/images/logo/ucla-logo.png" alt="UCLA Logo" class="org-logo" style="height: 60px;">
</a>
<a href="https://www.washington.edu/" target="_blank" rel="noopener noreferrer">
<img src="static/images/logo/uw-logo.png" alt="UW Logo" class="org-logo" style="height: 60px;">
</a>
<a href="https://www.hbku.edu.qa/en/qcri" target="_blank" rel="noopener noreferrer">
<img src="static/images/logo/QCRI-Logo.jpeg" alt="QCRI Logo" class="org-logo" style="height: 80px;">
</a>
<a href="https://research.google/" target="_blank" rel="noopener noreferrer">
<img src="static/images/logo/google_research-logo.png" alt="Google Research Logo" class="org-logo" style="height: 110px;">
</a>
<a href="https://www.stanford.edu/" target="_blank" rel="noopener noreferrer">
<img src="static/images/logo/stanford-logo.png" alt="Stanford Logo" class="org-logo" style="height: 55px;">
</a>
</div>
</div>
</div>
</div>
</section>
<footer class="footer">
<div class="container">
<div class="content has-text-centered">
<p>
This website is adapted from <a href="https://nerfies.github.io" target="_blank" rel="noopener noreferrer">Nerfies</a>, licensed under a
<a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/" target="_blank" rel="noopener noreferrer">Creative Commons Attribution-ShareAlike 4.0 International License</a>.
</p>
</div>
</div>
</footer>
<style>
.footer {
padding: 3rem 1.5rem;
margin-top: 2rem;
}
.footer a {
color: #485fc7;
text-decoration: underline;
}
.org-logos {
display: flex;
justify-content: center;
align-items: center;
flex-wrap: wrap;
gap: 2rem;
}
.org-logo {
height: 70px;
object-fit: contain;
margin: 0 2.5rem;
transition: transform 0.2s ease;
}
.org-logo:hover {
transform: scale(1.05);
}
@media screen and (max-width: 768px) {
.org-logo {
height: 50px;
margin: 1rem;
}
}
</style>
</body>
</html>