List your company

RAG Attack Defense on Natural Questions

0ASRTrustRAG
-3.821.8547.573.15May 1, 2026

Evaluation results

MethodLinks
TrustRAGAttack Scenario=GL2026.050——————0
CleanBaseAttack Scenario=GL2026.050——————0
CleanBase + MetaSecAlignAttack Scenario=GL2026.050——————0
TrustRAGAttack Scenario=PW2026.051——————1.6
TrustRAGAttack Scenario=PB2026.051——————1
PromptGuardAttack Scenario=FC2026.051——————0
CleanBase + MetaSecAlignAttack Scenario=NA2026.051——————0
TrustRAGAttack Scenario=NA2026.052——————87.2
TrustRAGAttack Scenario=CI2026.052——————2.4
TrustRAGAttack Scenario=FC2026.052——————2.2
TrustRAGAttack Scenario=CA2026.052——————3.2
TrustRAGAttack Scenario=MA2026.052——————2.6
PromptGuardAttack Scenario=CI2026.052——————0
PromptGuardAttack Scenario=CA2026.052——————0
CleanBaseAttack Scenario=CA2026.052——————0
PromptGuardAttack Scenario=NA2026.053——————0
PromptGuardAttack Scenario=MA2026.053——————0
CleanBaseAttack Scenario=NA2026.053——————0
CleanBase + MetaSecAlignAttack Scenario=CI2026.053——————1.2
CleanBase + MetaSecAlignAttack Scenario=CA2026.053——————0
CleanBaseAttack Scenario=FC2026.054——————1.6
PromptGuardAttack Scenario=PB2026.055——————3.6
CleanBaseAttack Scenario=CI2026.055——————1.2
CleanBase + MetaSecAlignAttack Scenario=FC2026.055——————1.6
CleanBase + MetaSecAlignAttack Scenario=MA2026.056——————3.4
TrustRAGAttack Scenario=PT2026.057——————1.4
PromptGuardAttack Scenario=GL2026.0510——————7
CleanBase + MetaSecAlignAttack Scenario=PB2026.0511——————9.2
CleanBaseAttack Scenario=MA2026.0512——————3.4
CleanBaseAttack Scenario=PB2026.0512——————9.2
PromptGuardAttack Scenario=PT2026.0517——————13.4
CleanBase + MetaSecAlignAttack Scenario=PT2026.0517——————16.6
CleanBaseAttack Scenario=PT2026.0520——————16.6
PromptGuardAttack Scenario=PW2026.0521——————9.2
DataSentinelAttack Scenario=CA2026.0531——————12.4
DataSentinelAttack Scenario=GL2026.0535——————42
MetaSecAlignAttack Scenario=FC2026.0539——————73.6
MetaSecAlignAttack Scenario=NA2026.0540——————94.8
MetaSecAlignAttack Scenario=CA2026.0543——————64.6
DataSentinelAttack Scenario=CI2026.0543——————27.2
DataSentinelAttack Scenario=FC2026.0543——————24.4
MetaSecAlignAttack Scenario=MA2026.0546——————72.8
MetaSecAlignAttack Scenario=GL2026.0550——————100
MetaSecAlignAttack Scenario=CI2026.0551——————80.8
DataSentinelAttack Scenario=MA2026.0553——————28
MetaSecAlignAttack Scenario=PT2026.0558——————97.4
No DefenseAttack Scenario=PT2026.0560——————97.4
CleanBase + MetaSecAlignAttack Scenario=PW2026.0560——————70.6
MetaSecAlignAttack Scenario=PB2026.0561——————98
DataSentinelAttack Scenario=PT2026.0561——————97
No DefenseAttack Scenario=GL2026.0570——————100
MetaSecAlignAttack Scenario=PW2026.0570——————97.4
CleanBaseAttack Scenario=PW2026.0573——————70.6
No DefenseAttack Scenario=CA2026.0574——————64.6
DataSentinelAttack Scenario=NA2026.0576——————77.6
No DefenseAttack Scenario=FC2026.0585——————73.6
DataSentinelAttack Scenario=PB2026.0586——————97.2
No DefenseAttack Scenario=PB2026.0588——————98
No DefenseAttack Scenario=MA2026.0589——————72.8
DataSentinelAttack Scenario=PW2026.0589——————96
No DefenseAttack Scenario=PW2026.0591——————97.4
No DefenseAttack Scenario=NA2026.0595——————94.8
No DefenseAttack Scenario=CI2026.0595——————80.8
L1-OnlyNumber of passages=500, Number of queries=200, Confidence Interval=95% Wilson CI2026.04—0080110—
L3-OnlyNumber of passages=500, Number of queries=200, Confidence Interval=95% Wilson CI2026.04—000000—
No DefenseNumber of passages=500, Number of queries=200, Confidence Interval=95% Wilson CI2026.04—13888.511——
RAGDefenderNumber of passages=500, Number of queries=200, Confidence Interval=95% Wilson CI2026.04—12.57.57.58.511——
RAGShieldNumber of passages=500, Number of queries=200, Confidence Interval=95% Wilson CI2026.04—000000—
RobustRAGNumber of passages=500, Number of queries=200, Confidence Interval=95% Wilson CI2026.04—0.50000——