Large Language Models (LLMs) have shown promise for empowering search agents to tackle complex information-seeking tasks. However, since acquiring external information necessitates web search, these agents are highly susceptible to various adversarial attacks (e.g., malicious information injection or stealthy context manipulation). Existing attack strategies inevitably rely on static and human-crafted templates, failing to emulate the dynamic and complex nature of real-world threats. In this paper, we propose an adversarial self-evolution framework that integrates bi-level optimization between the attacker's strategy evolution and the defender's adaptive mitigation, to expose the vulnerabilities of search agents. Specifically, we introduce the Contrastive Rollout Evolutionary Optimization (CREO) method to drive the directional evolution of the attack strategy population via contrastive evolutionary signals. To provide a stable environment for this continuous evolution, we further construct a controlled sandbox and design fine-grained metrics to quantify internal vulnerabilities. Extensive experiments across various multi-hop QA benchmarks and frontier LLMs demonstrate that the proposed framework yields attack efficacy superior to static baselines, underscoring the urgent need to develop more robust defense frameworks.
Figures & tables
Figure 1: Overview of the adversarial self-evolution framework. Left: Bi-level optimization drives attacker-defender co-evolution. Top Right: CREO drives directional mutation and crossover recombination of the attack strategy population via contrastive evolution signals. Bottom: The controlled sandbox ensures noise-free trace capture, while fine-grained metrics (Acc, ASR, APR) quantify the agent’s internal cognitive vulnerabilities.
Dataset
Model
Base
FactSwap
AuthChain
CREO Adversarial Self-Evolution (Ours, 3 Rounds)
Evo
Sync
Evo
Sync
Round 1 Evo
Round 1 Sync
Round 2 Evo
Round 2 Sync
Round 3 Evo
Round 3 Sync
Acc
ASR
Acc
ASR
Acc
ASR
Acc
ASR
Acc
ASR
Acc
ASR
Acc
ASR
Acc
ASR
Acc
ASR
Acc
ASR
HotpotQA
Qwen-Turbo
90
69
24
70
21
41
58
48
44
28
48
36
30
36
32
37
39
35
36
42
32
Qwen3-Max
92
71
08
77
08
49
47
69
25
20
20
22
12
16
11
23
08
22
11
25
15
DeepSeek-V3.2
88
67
08
72
07
37
55
56
25
11
24
11
07
11
13
21
10
09
06
13
11
Kimi-K2.5
93
72
06
89
01
52
34
68
18
45
08
50
01
46
04
48
03
47
05
38
03
Table 1: Main attack evaluation results (Acc and ASR) comparing static baselines against CREO’s three-round adversarial evolution under Evo and Sync conditions.
Dataset
Model
FactSwap
AuthChain
CREO (Ours)
Evo
Sync
Evo
Sync
R1E
R1S
R2E
R2S
R3E
R3S
HotpotQA
Qwen-Turbo
02
05
00
06
23
30
30
20
26
22
Qwen3-Max
12
10
04
06
59
65
73
68
67
60
DeepSeek-V3.2
18
14
08
17
65
82
75
69
81
75
Kimi-K2.5
15
07
04
03
29
38
34
33
30
45
GLM-5.1
02
04
02
02
57
13
16
17
14
12
Table 2: APR results highlighting the cognitive paralysis dimension induced by CREO’s evolved strategies across three rounds.
Variant
Evo.
Dir.
Qwen-T
Qwen3-M
Kimi
GLM
(HotpotQA)
(HotpotQA)
(2Wiki)
(2Wiki)
Static
×
×
38%
23%
31%
22%
Random
✓
×
34%
24%
36%
20%
Full
✓
✓
28%
20%
28%
16%
Table 3: Ablation Study (Round 1). We report True Accuracy (Acc ↓ ). Lower indicates stronger attack. Evo. : Evolution Action; Dir. : Directed Feedback.
Figure 2: Pairwise Jaccard similarity matrices of the final strategy pools (Qwen3-Max, HotpotQA). Full CREO maintains population diversity under the homogeneity constraint, while removing it leads to rapid redundant convergence across strategies.
Figure 3: Total interference rate under different defense modes (HotpotQA, 100 queries).
Figure 4: Round-by-round attack-defense trajectory on a representative query. Each round shows the evolved attack strategy alongside the agent’s reasoning outcome under the Evo and Sync conditions, respectively, illustrating how the two settings diverge over successive rounds.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Acc and ASR across adversarial evolution rounds (Part I): HotpotQA and 2WikiMultihopQA, shown for both Evo and Sync conditions.
Figure 6: Acc and ASR across adversarial evolution rounds (Part II): MuSiQue. APR comparison across all datasets and target models under CREO vs. static baselines.
Symbol
Meaning
Type or Data Format
q,D
Current question and multi-hop QA task domain
q is natural-language question text; D is the question domain.
Deval
Fixed fitness evaluation set
A fixed subset of training questions.
t,K,N
Round index, initialization rollout count, and pool capacity
t is a nonnegative integer; K and N are positive integers.
P0,Pt
Initial and round- t strategy pools
Sets of strategy individuals, with at most N individuals per pool.
Kt
Round- t defensive knowledge
Structured system-prompt text derived from the strategy pool disclosed under the corresponding evaluation condition.
s,sparent,schild
Strategy individual, parent, and child
Triplets containing a mechanism, fitness, and contextual profile.
Appendix
Table 4: Variables and data representations used in the framework.
Function or Operation
Input
Output and Format
AP(q)
Question and strategy-pool context; document generation also uses the reference answer.
Generated adversarial documents and their retrieval injection. The generation template returns JSON with text , url , and title .
AK(⋅)
Question, retrieval results, and defensive-knowledge context.
The agent’s interaction trace and final answer.
J(⋅)
Complete interaction trace, with the question, ground truth, and forged document as context.
A cognitive-state label in {DA,CP,Refuted,Neutral} .
Acc(⋅)
Answer correctness and cognitive state.
A binary indicator of a correct answer without DA or CP; dataset accuracy is its mean.
I(⋅)
A Boolean condition.
1 if the condition holds, and 0 otherwise.
G(s)
Attack outcomes for strategy s on the fixed evaluation set.
The empirical proportion of DA or CP outcomes, in [0,1] .
Appendix
Table 5: Inputs, outputs, and data formats of the framework’s functions.
Annotator
Agreement
Cohen’s κ
Flagged Samples
Human 1
96/100
0.947
3, 24, 69, 75
Human 2
97/100
0.960
11, 24, 75
GPT-5.6 Sol
97/100
0.960
10, 11, 91
Claude fable-5
97/100
0.960
11, 75, 91
Appendix
Table 6: Agreement of two human annotators and two LLMs with the original judge on 100 stratified reasoning traces.
Method
ASR Evo
ASR Sync
AutoDAN-Turbo
14
9
PromptBreeder
16
9
Breaking the Ceiling
15
4
CREO (Ours)
20
12
Appendix
Table 7: ASR (%) of baseline algorithms transferred into our framework compared with CREO on HotpotQA, with Qwen3-Max as the target model.
Dataset
Model
With ext.
Without ext.
Diff. (pp)
HotpotQA
Qwen3-Max
92
74
−18
HotpotQA
GLM-5.1
92
70
−22
2WikiMultihopQA
Qwen3-Max
79
52
−27
2WikiMultihopQA
GLM-5.1
74
53
−21
MuSiQue
Qwen3-Max
90
25
−65
MuSiQue
GLM-5.1
82
30
−52
Appendix
Table 8: Answer accuracy (%) of two models with and without external knowledge. The difference is the accuracy without external knowledge minus that with external knowledge, in percentage points.
Dataset
Model
R1
R2
R3
HotpotQA
Qwen-Turbo
30
24
20
HotpotQA
Qwen3-Max
25
20
15
2WikiMultihopQA
Qwen-Turbo
15
15
10
2WikiMultihopQA
Qwen3-Max
8
7
3
MuSiQue
Qwen-Turbo
21
20
12
MuSiQue
Qwen3-Max
17
16
11
Appendix
Table 9: Acc (%) across three evolution rounds under fixed defensive knowledge derived from the Round 0 strategy pool.
Model
R1
R2
R3
Qwen3-Max
20
18
18
GLM-5.1
30
27
28
Appendix
Table 10: Acc (%) of the Round 1, 2, and 3 strategy pools evaluated without defense on HotpotQA.
Attacker
D1
D2
D3
A1
22
20
28
A2
19
21
22
A3
21
18
23
Appendix
Table 11: Acc (%) for cross-round attacker–defender combinations on HotpotQA, with Qwen3-Max as the target model.
Dataset
Model
Wall Time
API Calls
Prompt Tokens
Completion Tokens
Total Tokens
HotpotQA
Qwen-Turbo
22h 52m
14,279
43,894,767
1,323,069
45,217,836
DeepSeek-V3.2
51h 14m
37,697
236,485,091
4,386,538
240,871,629
Qwen3-Max
21h 07m
21,164
92,436,744
1,889,224
94,325,968
Kimi-K2.5
13h 33m
13,583
63,398,393
1,190,712
64,589,105
GLM-5.1
22h 25m
20,844
92,250,803
3,194,661
95,445,464
2WikiMultihopQA
Qwen-Turbo
30h 55m
14,310
35,045,955
1,357,964
36,403,919
Appendix
Table 12: Computational cost per dataset and target model.
Method
Tokens per query
FactSwap
874
AuthChain
1,841
CREO
2,690
Appendix
Table 13: Per-query token cost of generating an attack document on HotpotQA, with Qwen3-Max as the target model.
School of Cyber Science and Technology, University of Science and Technology of China · College of Computing and Data Science, Nanyang Technological University