白泽(Whitzard)女娲实验室 ↗
NUWANUWA INDEX

搜索 NUWA Index

研究 · 测量协议 · 评测结果 · 观点

English

RESEARCH RECORD

Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents

Hanrong Zhang · Jingyuan Huang · Kai Mei · Yifei Yao · Zhenting Wang · Chenlu Zhan · Hongwei Wang · Yongfeng Zhang · Zhejiang University; Rutgers University

原始文献 ↗ · BibTeX · RIS

本次捕获版本报告多个接口存在脆弱性,所测防御保护有限;安全性需要与任务效用共同检查。

研究问题

智能体的哪些接口暴露出不同的攻击与防御失效模式?

设计与前提

ASB 区分系统与用户提示、工具响应和检索记忆,在共同智能体框架中评测攻击与防御。

研究发现

本次捕获版本报告多个接口存在脆弱性,所测防御保护有限;安全性需要与任务效用共同检查。

局限与反证

攻击面、基础模型和防御配置不同;汇总攻击率不是部署风险概率。方法与指标数量随版本变化,不能沿用旧摘要数字。

来源证据

系统测量原值条件与位置
Gemma2-9BAttack success rate87.10%%DPI#S5.T5; row Gemma2-9B; condition DPI; column Attack success rate
Gemma2-9BRefusal rate4.30%%DPI#S5.T5; row Gemma2-9B; condition DPI; column Refusal rate
Gemma2-9BAttack success rate14.20%%IPI#S5.T5; row Gemma2-9B; condition IPI; column Attack success rate
Gemma2-9BRefusal rate15.00%%IPI#S5.T5; row Gemma2-9B; condition IPI; column Refusal rate
Gemma2-9BAttack success rate6.85%%Memory poisoning#S5.T5; row Gemma2-9B; condition Memory poisoning; column Attack success rate
Gemma2-9BRefusal rate9.85%%Memory poisoning#S5.T5; row Gemma2-9B; condition Memory poisoning; column Refusal rate
Gemma2-9BAttack success rate92.17%%Mixed attack#S5.T5; row Gemma2-9B; condition Mixed attack; column Attack success rate
Gemma2-9BRefusal rate1.33%%Mixed attack#S5.T5; row Gemma2-9B; condition Mixed attack; column Refusal rate
Gemma2-9BAttack success rate39.75%%PoT backdoor#S5.T5; row Gemma2-9B; condition PoT backdoor; column Attack success rate
Gemma2-9BRefusal rate5.25%%PoT backdoor#S5.T5; row Gemma2-9B; condition PoT backdoor; column Refusal rate
Gemma2-9BAttack success rate48.01%%Average across attack types#S5.T5; row Gemma2-9B; condition Average across attack types; column Attack success rate
Gemma2-9BRefusal rate7.15%%Average across attack types#S5.T5; row Gemma2-9B; condition Average across attack types; column Refusal rate
Gemma2-27BAttack success rate96.75%%DPI#S5.T5; row Gemma2-27B; condition DPI; column Attack success rate
Gemma2-27BRefusal rate0.90%%DPI#S5.T5; row Gemma2-27B; condition DPI; column Refusal rate
Gemma2-27BAttack success rate14.20%%IPI#S5.T5; row Gemma2-27B; condition IPI; column Attack success rate
Gemma2-27BRefusal rate3.90%%IPI#S5.T5; row Gemma2-27B; condition IPI; column Refusal rate
Gemma2-27BAttack success rate6.25%%Memory poisoning#S5.T5; row Gemma2-27B; condition Memory poisoning; column Attack success rate
Gemma2-27BRefusal rate5.45%%Memory poisoning#S5.T5; row Gemma2-27B; condition Memory poisoning; column Refusal rate
Gemma2-27BAttack success rate100.00%%Mixed attack#S5.T5; row Gemma2-27B; condition Mixed attack; column Attack success rate
Gemma2-27BRefusal rate0.50%%Mixed attack#S5.T5; row Gemma2-27B; condition Mixed attack; column Refusal rate
Gemma2-27BAttack success rate54.50%%PoT backdoor#S5.T5; row Gemma2-27B; condition PoT backdoor; column Attack success rate
Gemma2-27BRefusal rate3.50%%PoT backdoor#S5.T5; row Gemma2-27B; condition PoT backdoor; column Refusal rate
Gemma2-27BAttack success rate54.34%%Average across attack types#S5.T5; row Gemma2-27B; condition Average across attack types; column Attack success rate
Gemma2-27BRefusal rate2.85%%Average across attack types#S5.T5; row Gemma2-27B; condition Average across attack types; column Refusal rate
LLaMA3-8BAttack success rate25.20%%DPI#S5.T5; row LLaMA3-8B; condition DPI; column Attack success rate
LLaMA3-8BRefusal rate7.45%%DPI#S5.T5; row LLaMA3-8B; condition DPI; column Refusal rate
LLaMA3-8BAttack success rate10.55%%IPI#S5.T5; row LLaMA3-8B; condition IPI; column Attack success rate
LLaMA3-8BRefusal rate3.00%%IPI#S5.T5; row LLaMA3-8B; condition IPI; column Refusal rate
LLaMA3-8BAttack success rate3.30%%Memory poisoning#S5.T5; row LLaMA3-8B; condition Memory poisoning; column Attack success rate
LLaMA3-8BRefusal rate5.45%%Memory poisoning#S5.T5; row LLaMA3-8B; condition Memory poisoning; column Refusal rate
LLaMA3-8BAttack success rate40.75%%Mixed attack#S5.T5; row LLaMA3-8B; condition Mixed attack; column Attack success rate
LLaMA3-8BRefusal rate5.75%%Mixed attack#S5.T5; row LLaMA3-8B; condition Mixed attack; column Refusal rate
LLaMA3-8BAttack success rate21.50%%PoT backdoor#S5.T5; row LLaMA3-8B; condition PoT backdoor; column Attack success rate
LLaMA3-8BRefusal rate2.50%%PoT backdoor#S5.T5; row LLaMA3-8B; condition PoT backdoor; column Refusal rate
LLaMA3-8BAttack success rate20.26%%Average across attack types#S5.T5; row LLaMA3-8B; condition Average across attack types; column Attack success rate
LLaMA3-8BRefusal rate4.83%%Average across attack types#S5.T5; row LLaMA3-8B; condition Average across attack types; column Refusal rate
LLaMA3-70BAttack success rate86.15%%DPI#S5.T5; row LLaMA3-70B; condition DPI; column Attack success rate
LLaMA3-70BRefusal rate7.80%%DPI#S5.T5; row LLaMA3-70B; condition DPI; column Refusal rate
LLaMA3-70BAttack success rate43.70%%IPI#S5.T5; row LLaMA3-70B; condition IPI; column Attack success rate
LLaMA3-70BRefusal rate3.00%%IPI#S5.T5; row LLaMA3-70B; condition IPI; column Refusal rate
LLaMA3-70BAttack success rate1.85%%Memory poisoning#S5.T5; row LLaMA3-70B; condition Memory poisoning; column Attack success rate
LLaMA3-70BRefusal rate1.80%%Memory poisoning#S5.T5; row LLaMA3-70B; condition Memory poisoning; column Refusal rate
LLaMA3-70BAttack success rate85.50%%Mixed attack#S5.T5; row LLaMA3-70B; condition Mixed attack; column Attack success rate
LLaMA3-70BRefusal rate6.50%%Mixed attack#S5.T5; row LLaMA3-70B; condition Mixed attack; column Refusal rate
LLaMA3-70BAttack success rate57.00%%PoT backdoor#S5.T5; row LLaMA3-70B; condition PoT backdoor; column Attack success rate
LLaMA3-70BRefusal rate2.00%%PoT backdoor#S5.T5; row LLaMA3-70B; condition PoT backdoor; column Refusal rate
LLaMA3-70BAttack success rate54.84%%Average across attack types#S5.T5; row LLaMA3-70B; condition Average across attack types; column Attack success rate
LLaMA3-70BRefusal rate4.22%%Average across attack types#S5.T5; row LLaMA3-70B; condition Average across attack types; column Refusal rate
LLaMA3.1-8BAttack success rate51.10%%DPI#S5.T5; row LLaMA3.1-8B; condition DPI; column Attack success rate
LLaMA3.1-8BRefusal rate5.20%%DPI#S5.T5; row LLaMA3.1-8B; condition DPI; column Refusal rate
LLaMA3.1-8BAttack success rate6.40%%IPI#S5.T5; row LLaMA3.1-8B; condition IPI; column Attack success rate
LLaMA3.1-8BRefusal rate1.85%%IPI#S5.T5; row LLaMA3.1-8B; condition IPI; column Refusal rate
LLaMA3.1-8BAttack success rate25.65%%Memory poisoning#S5.T5; row LLaMA3.1-8B; condition Memory poisoning; column Attack success rate
LLaMA3.1-8BRefusal rate6.75%%Memory poisoning#S5.T5; row LLaMA3.1-8B; condition Memory poisoning; column Refusal rate
LLaMA3.1-8BAttack success rate73.50%%Mixed attack#S5.T5; row LLaMA3.1-8B; condition Mixed attack; column Attack success rate
LLaMA3.1-8BRefusal rate3.50%%Mixed attack#S5.T5; row LLaMA3.1-8B; condition Mixed attack; column Refusal rate
LLaMA3.1-8BAttack success rate19.00%%PoT backdoor#S5.T5; row LLaMA3.1-8B; condition PoT backdoor; column Attack success rate
LLaMA3.1-8BRefusal rate5.75%%PoT backdoor#S5.T5; row LLaMA3.1-8B; condition PoT backdoor; column Refusal rate
LLaMA3.1-8BAttack success rate35.13%%Average across attack types#S5.T5; row LLaMA3.1-8B; condition Average across attack types; column Attack success rate
LLaMA3.1-8BRefusal rate4.61%%Average across attack types#S5.T5; row LLaMA3.1-8B; condition Average across attack types; column Refusal rate
LLaMA3.1-70BAttack success rate85.65%%DPI#S5.T5; row LLaMA3.1-70B; condition DPI; column Attack success rate
LLaMA3.1-70BRefusal rate5.30%%DPI#S5.T5; row LLaMA3.1-70B; condition DPI; column Refusal rate
LLaMA3.1-70BAttack success rate12.10%%IPI#S5.T5; row LLaMA3.1-70B; condition IPI; column Attack success rate
LLaMA3.1-70BRefusal rate4.95%%IPI#S5.T5; row LLaMA3.1-70B; condition IPI; column Refusal rate
LLaMA3.1-70BAttack success rate2.85%%Memory poisoning#S5.T5; row LLaMA3.1-70B; condition Memory poisoning; column Attack success rate
LLaMA3.1-70BRefusal rate2.20%%Memory poisoning#S5.T5; row LLaMA3.1-70B; condition Memory poisoning; column Refusal rate
LLaMA3.1-70BAttack success rate94.50%%Mixed attack#S5.T5; row LLaMA3.1-70B; condition Mixed attack; column Attack success rate
LLaMA3.1-70BRefusal rate1.25%%Mixed attack#S5.T5; row LLaMA3.1-70B; condition Mixed attack; column Refusal rate
LLaMA3.1-70BAttack success rate59.75%%PoT backdoor#S5.T5; row LLaMA3.1-70B; condition PoT backdoor; column Attack success rate
LLaMA3.1-70BRefusal rate6.25%%PoT backdoor#S5.T5; row LLaMA3.1-70B; condition PoT backdoor; column Refusal rate
LLaMA3.1-70BAttack success rate50.97%%Average across attack types#S5.T5; row LLaMA3.1-70B; condition Average across attack types; column Attack success rate
LLaMA3.1-70BRefusal rate3.99%%Average across attack types#S5.T5; row LLaMA3.1-70B; condition Average across attack types; column Refusal rate
Mixtral-8x7BAttack success rate25.85%%DPI#S5.T5; row Mixtral-8x7B; condition DPI; column Attack success rate
Mixtral-8x7BRefusal rate9.55%%DPI#S5.T5; row Mixtral-8x7B; condition DPI; column Refusal rate
Mixtral-8x7BAttack success rate4.80%%IPI#S5.T5; row Mixtral-8x7B; condition IPI; column Attack success rate
Mixtral-8x7BRefusal rate8.55%%IPI#S5.T5; row Mixtral-8x7B; condition IPI; column Refusal rate
Mixtral-8x7BAttack success rate4.90%%Memory poisoning#S5.T5; row Mixtral-8x7B; condition Memory poisoning; column Attack success rate
Mixtral-8x7BRefusal rate1.35%%Memory poisoning#S5.T5; row Mixtral-8x7B; condition Memory poisoning; column Refusal rate
Mixtral-8x7BAttack success rate54.75%%Mixed attack#S5.T5; row Mixtral-8x7B; condition Mixed attack; column Attack success rate
Mixtral-8x7BRefusal rate6.75%%Mixed attack#S5.T5; row Mixtral-8x7B; condition Mixed attack; column Refusal rate
Mixtral-8x7BAttack success rate4.75%%PoT backdoor#S5.T5; row Mixtral-8x7B; condition PoT backdoor; column Attack success rate
Mixtral-8x7BRefusal rate13.25%%PoT backdoor#S5.T5; row Mixtral-8x7B; condition PoT backdoor; column Refusal rate
Mixtral-8x7BAttack success rate19.01%%Average across attack types#S5.T5; row Mixtral-8x7B; condition Average across attack types; column Attack success rate
Mixtral-8x7BRefusal rate7.89%%Average across attack types#S5.T5; row Mixtral-8x7B; condition Average across attack types; column Refusal rate
Qwen2-7BAttack success rate55.20%%DPI#S5.T5; row Qwen2-7B; condition DPI; column Attack success rate
Qwen2-7BRefusal rate7.70%%DPI#S5.T5; row Qwen2-7B; condition DPI; column Refusal rate
Qwen2-7BAttack success rate9.00%%IPI#S5.T5; row Qwen2-7B; condition IPI; column Attack success rate
Qwen2-7BRefusal rate6.00%%IPI#S5.T5; row Qwen2-7B; condition IPI; column Refusal rate
Qwen2-7BAttack success rate2.85%%Memory poisoning#S5.T5; row Qwen2-7B; condition Memory poisoning; column Attack success rate
Qwen2-7BRefusal rate4.95%%Memory poisoning#S5.T5; row Qwen2-7B; condition Memory poisoning; column Refusal rate
Qwen2-7BAttack success rate76.00%%Mixed attack#S5.T5; row Qwen2-7B; condition Mixed attack; column Attack success rate
Qwen2-7BRefusal rate2.50%%Mixed attack#S5.T5; row Qwen2-7B; condition Mixed attack; column Refusal rate
Qwen2-7BAttack success rate12.25%%PoT backdoor#S5.T5; row Qwen2-7B; condition PoT backdoor; column Attack success rate
Qwen2-7BRefusal rate4.50%%PoT backdoor#S5.T5; row Qwen2-7B; condition PoT backdoor; column Refusal rate
Qwen2-7BAttack success rate31.06%%Average across attack types#S5.T5; row Qwen2-7B; condition Average across attack types; column Attack success rate
Qwen2-7BRefusal rate5.13%%Average across attack types#S5.T5; row Qwen2-7B; condition Average across attack types; column Refusal rate
Qwen2-72BAttack success rate86.95%%DPI#S5.T5; row Qwen2-72B; condition DPI; column Attack success rate
Qwen2-72BRefusal rate4.20%%DPI#S5.T5; row Qwen2-72B; condition DPI; column Refusal rate
Qwen2-72BAttack success rate21.35%%IPI#S5.T5; row Qwen2-72B; condition IPI; column Attack success rate
Qwen2-72BRefusal rate16.55%%IPI#S5.T5; row Qwen2-72B; condition IPI; column Refusal rate
Qwen2-72BAttack success rate3.95%%Memory poisoning#S5.T5; row Qwen2-72B; condition Memory poisoning; column Attack success rate
Qwen2-72BRefusal rate5.45%%Memory poisoning#S5.T5; row Qwen2-72B; condition Memory poisoning; column Refusal rate
Qwen2-72BAttack success rate98.50%%Mixed attack#S5.T5; row Qwen2-72B; condition Mixed attack; column Attack success rate
Qwen2-72BRefusal rate0.75%%Mixed attack#S5.T5; row Qwen2-72B; condition Mixed attack; column Refusal rate
Qwen2-72BAttack success rate57.75%%PoT backdoor#S5.T5; row Qwen2-72B; condition PoT backdoor; column Attack success rate
Qwen2-72BRefusal rate4.75%%PoT backdoor#S5.T5; row Qwen2-72B; condition PoT backdoor; column Refusal rate
Qwen2-72BAttack success rate53.70%%Average across attack types#S5.T5; row Qwen2-72B; condition Average across attack types; column Attack success rate
Qwen2-72BRefusal rate6.34%%Average across attack types#S5.T5; row Qwen2-72B; condition Average across attack types; column Refusal rate
Claude3.5 SonnetAttack success rate90.75%%DPI#S5.T5; row Claude3.5 Sonnet; condition DPI; column Attack success rate
Claude3.5 SonnetRefusal rate7.65%%DPI#S5.T5; row Claude3.5 Sonnet; condition DPI; column Refusal rate
Claude3.5 SonnetAttack success rate59.70%%IPI#S5.T5; row Claude3.5 Sonnet; condition IPI; column Attack success rate
Claude3.5 SonnetRefusal rate25.50%%IPI#S5.T5; row Claude3.5 Sonnet; condition IPI; column Refusal rate
Claude3.5 SonnetAttack success rate19.75%%Memory poisoning#S5.T5; row Claude3.5 Sonnet; condition Memory poisoning; column Attack success rate
Claude3.5 SonnetRefusal rate1.20%%Memory poisoning#S5.T5; row Claude3.5 Sonnet; condition Memory poisoning; column Refusal rate
Claude3.5 SonnetAttack success rate94.50%%Mixed attack#S5.T5; row Claude3.5 Sonnet; condition Mixed attack; column Attack success rate
Claude3.5 SonnetRefusal rate6.25%%Mixed attack#S5.T5; row Claude3.5 Sonnet; condition Mixed attack; column Refusal rate
Claude3.5 SonnetAttack success rate17.50%%PoT backdoor#S5.T5; row Claude3.5 Sonnet; condition PoT backdoor; column Attack success rate
Claude3.5 SonnetRefusal rate11.75%%PoT backdoor#S5.T5; row Claude3.5 Sonnet; condition PoT backdoor; column Refusal rate
Claude3.5 SonnetAttack success rate56.44%%Average across attack types#S5.T5; row Claude3.5 Sonnet; condition Average across attack types; column Attack success rate
Claude3.5 SonnetRefusal rate10.47%%Average across attack types#S5.T5; row Claude3.5 Sonnet; condition Average across attack types; column Refusal rate
GPT-3.5 TurboAttack success rate98.40%%DPI#S5.T5; row GPT-3.5 Turbo; condition DPI; column Attack success rate
GPT-3.5 TurboRefusal rate3.00%%DPI#S5.T5; row GPT-3.5 Turbo; condition DPI; column Refusal rate
GPT-3.5 TurboAttack success rate55.10%%IPI#S5.T5; row GPT-3.5 Turbo; condition IPI; column Attack success rate
GPT-3.5 TurboRefusal rate16.85%%IPI#S5.T5; row GPT-3.5 Turbo; condition IPI; column Refusal rate
GPT-3.5 TurboAttack success rate9.30%%Memory poisoning#S5.T5; row GPT-3.5 Turbo; condition Memory poisoning; column Attack success rate
GPT-3.5 TurboRefusal rate0.30%%Memory poisoning#S5.T5; row GPT-3.5 Turbo; condition Memory poisoning; column Refusal rate
GPT-3.5 TurboAttack success rate99.75%%Mixed attack#S5.T5; row GPT-3.5 Turbo; condition Mixed attack; column Attack success rate
GPT-3.5 TurboRefusal rate0.00%%Mixed attack#S5.T5; row GPT-3.5 Turbo; condition Mixed attack; column Refusal rate
GPT-3.5 TurboAttack success rate8.25%%PoT backdoor#S5.T5; row GPT-3.5 Turbo; condition PoT backdoor; column Attack success rate
GPT-3.5 TurboRefusal rate10.75%%PoT backdoor#S5.T5; row GPT-3.5 Turbo; condition PoT backdoor; column Refusal rate
GPT-3.5 TurboAttack success rate54.16%%Average across attack types#S5.T5; row GPT-3.5 Turbo; condition Average across attack types; column Attack success rate
GPT-3.5 TurboRefusal rate6.18%%Average across attack types#S5.T5; row GPT-3.5 Turbo; condition Average across attack types; column Refusal rate
GPT-4oAttack success rate60.35%%DPI#S5.T5; row GPT-4o; condition DPI; column Attack success rate
GPT-4oRefusal rate20.05%%DPI#S5.T5; row GPT-4o; condition DPI; column Refusal rate
GPT-4oAttack success rate62.45%%IPI#S5.T5; row GPT-4o; condition IPI; column Attack success rate
GPT-4oRefusal rate6.50%%IPI#S5.T5; row GPT-4o; condition IPI; column Refusal rate
GPT-4oAttack success rate10.00%%Memory poisoning#S5.T5; row GPT-4o; condition Memory poisoning; column Attack success rate
GPT-4oRefusal rate11.75%%Memory poisoning#S5.T5; row GPT-4o; condition Memory poisoning; column Refusal rate
GPT-4oAttack success rate89.25%%Mixed attack#S5.T5; row GPT-4o; condition Mixed attack; column Attack success rate
GPT-4oRefusal rate5.50%%Mixed attack#S5.T5; row GPT-4o; condition Mixed attack; column Refusal rate
GPT-4oAttack success rate100.00%%PoT backdoor#S5.T5; row GPT-4o; condition PoT backdoor; column Attack success rate
GPT-4oRefusal rate0.25%%PoT backdoor#S5.T5; row GPT-4o; condition PoT backdoor; column Refusal rate
GPT-4oAttack success rate64.41%%Average across attack types#S5.T5; row GPT-4o; condition Average across attack types; column Attack success rate
GPT-4oRefusal rate8.81%%Average across attack types#S5.T5; row GPT-4o; condition Average across attack types; column Refusal rate
GPT-4o-miniAttack success rate95.45%%DPI#S5.T5; row GPT-4o-mini; condition DPI; column Attack success rate
GPT-4o-miniRefusal rate1.85%%DPI#S5.T5; row GPT-4o-mini; condition DPI; column Refusal rate
GPT-4o-miniAttack success rate44.55%%IPI#S5.T5; row GPT-4o-mini; condition IPI; column Attack success rate
GPT-4o-miniRefusal rate0.25%%IPI#S5.T5; row GPT-4o-mini; condition IPI; column Refusal rate
GPT-4o-miniAttack success rate5.50%%Memory poisoning#S5.T5; row GPT-4o-mini; condition Memory poisoning; column Attack success rate
GPT-4o-miniRefusal rate3.65%%Memory poisoning#S5.T5; row GPT-4o-mini; condition Memory poisoning; column Refusal rate
GPT-4o-miniAttack success rate96.75%%Mixed attack#S5.T5; row GPT-4o-mini; condition Mixed attack; column Attack success rate
GPT-4o-miniRefusal rate1.25%%Mixed attack#S5.T5; row GPT-4o-mini; condition Mixed attack; column Refusal rate
GPT-4o-miniAttack success rate95.50%%PoT backdoor#S5.T5; row GPT-4o-mini; condition PoT backdoor; column Attack success rate
GPT-4o-miniRefusal rate0.00%%PoT backdoor#S5.T5; row GPT-4o-mini; condition PoT backdoor; column Refusal rate
GPT-4o-miniAttack success rate67.55%%Average across attack types#S5.T5; row GPT-4o-mini; condition Average across attack types; column Attack success rate
GPT-4o-miniRefusal rate1.40%%Average across attack types#S5.T5; row GPT-4o-mini; condition Average across attack types; column Refusal rate
AverageAttack success rate72.68%%DPI#S5.T5; row Average; condition DPI; column Attack success rate
AverageRefusal rate6.53%%DPI#S5.T5; row Average; condition DPI; column Refusal rate
AverageAttack success rate27.55%%IPI#S5.T5; row Average; condition IPI; column Attack success rate
AverageRefusal rate8.61%%IPI#S5.T5; row Average; condition IPI; column Refusal rate
AverageAttack success rate7.92%%Memory poisoning#S5.T5; row Average; condition Memory poisoning; column Attack success rate
AverageRefusal rate4.63%%Memory poisoning#S5.T5; row Average; condition Memory poisoning; column Refusal rate
AverageAttack success rate84.30%%Mixed attack#S5.T5; row Average; condition Mixed attack; column Attack success rate
AverageRefusal rate3.22%%Mixed attack#S5.T5; row Average; condition Mixed attack; column Refusal rate
AverageAttack success rate42.12%%PoT backdoor#S5.T5; row Average; condition PoT backdoor; column Attack success rate
AverageRefusal rate5.42%%PoT backdoor#S5.T5; row Average; condition PoT backdoor; column Refusal rate
AverageAttack success rate46.91%%Average across attack types#S5.T5; row Average; condition Average across attack types; column Attack success rate
AverageRefusal rate5.68%%Average across attack types#S5.T5; row Average; condition Average across attack types; column Refusal rate
Claude-3.5 SonnetPerformance under no attack100.00%No attack#S5.T6; row Claude-3.5 Sonnet; condition No attack; column Performance under no attack
Claude-3.5 SonnetAttack success rate56.44%Average across attack types#S5.T6; row Claude-3.5 Sonnet; condition Average across attack types; column Attack success rate
Claude-3.5 SonnetNRP43.56%Normalized risk-adjusted performance#S5.T6; row Claude-3.5 Sonnet; condition Normalized risk-adjusted performance; column NRP
LLaMA3-70BPerformance under no attack66.50%No attack#S5.T6; row LLaMA3-70B; condition No attack; column Performance under no attack
LLaMA3-70BAttack success rate54.84%Average across attack types#S5.T6; row LLaMA3-70B; condition Average across attack types; column Attack success rate
LLaMA3-70BNRP30.03%Normalized risk-adjusted performance#S5.T6; row LLaMA3-70B; condition Normalized risk-adjusted performance; column NRP
GPT-4oPerformance under no attack79.00%No attack#S5.T6; row GPT-4o; condition No attack; column Performance under no attack
GPT-4oAttack success rate64.41%Average across attack types#S5.T6; row GPT-4o; condition Average across attack types; column Attack success rate
GPT-4oNRP28.12%Normalized risk-adjusted performance#S5.T6; row GPT-4o; condition Normalized risk-adjusted performance; column NRP
GPT-4o-miniPerformance under no attack50.00%No attack#S5.T6; row GPT-4o-mini; condition No attack; column Performance under no attack
GPT-4o-miniAttack success rate67.55%Average across attack types#S5.T6; row GPT-4o-mini; condition Average across attack types; column Attack success rate
GPT-4o-miniNRP16.23%Normalized risk-adjusted performance#S5.T6; row GPT-4o-mini; condition Normalized risk-adjusted performance; column NRP
Gemma2-27BPerformance under no attack31.50%No attack#S5.T6; row Gemma2-27B; condition No attack; column Performance under no attack
Gemma2-27BAttack success rate54.34%Average across attack types#S5.T6; row Gemma2-27B; condition Average across attack types; column Attack success rate
Gemma2-27BNRP14.38%Normalized risk-adjusted performance#S5.T6; row Gemma2-27B; condition Normalized risk-adjusted performance; column NRP
LLaMA3.1-70BPerformance under no attack21.25%No attack#S5.T6; row LLaMA3.1-70B; condition No attack; column Performance under no attack
LLaMA3.1-70BAttack success rate50.97%Average across attack types#S5.T6; row LLaMA3.1-70B; condition Average across attack types; column Attack success rate
LLaMA3.1-70BNRP10.42%Normalized risk-adjusted performance#S5.T6; row LLaMA3.1-70B; condition Normalized risk-adjusted performance; column NRP
Qwen2-7BPerformance under no attack9.75%No attack#S5.T6; row Qwen2-7B; condition No attack; column Performance under no attack
Qwen2-7BAttack success rate31.06%Average across attack types#S5.T6; row Qwen2-7B; condition Average across attack types; column Attack success rate
Qwen2-7BNRP6.72%Normalized risk-adjusted performance#S5.T6; row Qwen2-7B; condition Normalized risk-adjusted performance; column NRP
Gemma2-9BPerformance under no attack10.75%No attack#S5.T6; row Gemma2-9B; condition No attack; column Performance under no attack
Gemma2-9BAttack success rate48.01%Average across attack types#S5.T6; row Gemma2-9B; condition Average across attack types; column Attack success rate
Gemma2-9BNRP5.59%Normalized risk-adjusted performance#S5.T6; row Gemma2-9B; condition Normalized risk-adjusted performance; column NRP
GPT-3.5 TurboPerformance under no attack8.00%No attack#S5.T6; row GPT-3.5 Turbo; condition No attack; column Performance under no attack
GPT-3.5 TurboAttack success rate54.16%Average across attack types#S5.T6; row GPT-3.5 Turbo; condition Average across attack types; column Attack success rate
GPT-3.5 TurboNRP3.67%Normalized risk-adjusted performance#S5.T6; row GPT-3.5 Turbo; condition Normalized risk-adjusted performance; column NRP
Qwen2-72BPerformance under no attack4.00%No attack#S5.T6; row Qwen2-72B; condition No attack; column Performance under no attack
Qwen2-72BAttack success rate53.70%Average across attack types#S5.T6; row Qwen2-72B; condition Average across attack types; column Attack success rate
Qwen2-72BNRP1.85%Normalized risk-adjusted performance#S5.T6; row Qwen2-72B; condition Normalized risk-adjusted performance; column NRP
LLaMA3-8BPerformance under no attack1.50%No attack#S5.T6; row LLaMA3-8B; condition No attack; column Performance under no attack
LLaMA3-8BAttack success rate20.26%Average across attack types#S5.T6; row LLaMA3-8B; condition Average across attack types; column Attack success rate
LLaMA3-8BNRP1.20%Normalized risk-adjusted performance#S5.T6; row LLaMA3-8B; condition Normalized risk-adjusted performance; column NRP
LLaMA3.1-8BPerformance under no attack0.75%No attack#S5.T6; row LLaMA3.1-8B; condition No attack; column Performance under no attack
LLaMA3.1-8BAttack success rate35.13%Average across attack types#S5.T6; row LLaMA3.1-8B; condition Average across attack types; column Attack success rate
LLaMA3.1-8BNRP0.49%Normalized risk-adjusted performance#S5.T6; row LLaMA3.1-8B; condition Normalized risk-adjusted performance; column NRP
Mixtral-8x7BPerformance under no attack0.00%No attack#S5.T6; row Mixtral-8x7B; condition No attack; column Performance under no attack
Mixtral-8x7BAttack success rate19.01%Average across attack types#S5.T6; row Mixtral-8x7B; condition Average across attack types; column Attack success rate
Mixtral-8x7BNRP0.00%Normalized risk-adjusted performance#S5.T6; row Mixtral-8x7B; condition Normalized risk-adjusted performance; column NRP
AveragePerformance under no attack29.46%No attack#S5.T6; row Average; condition No attack; column Performance under no attack
AverageAttack success rate46.91%Average across attack types#S5.T6; row Average; condition Average across attack types; column Attack success rate
AverageNRP15.64%Normalized risk-adjusted performance#S5.T6; row Average; condition Normalized risk-adjusted performance; column NRP
Gemma2-9BAttack success rate91.00%%DPI baseline#S5.T8; row Gemma2-9B; condition DPI baseline; column Attack success rate
Gemma2-9BAttack success rate91.75%%Delimiter#S5.T8; row Gemma2-9B; condition Delimiter; column Attack success rate
Gemma2-9BAttack success rate62.50%%Paraphrase#S5.T8; row Gemma2-9B; condition Paraphrase; column Attack success rate
Gemma2-9BAttack success rate91.00%%Instructional#S5.T8; row Gemma2-9B; condition Instructional; column Attack success rate
Gemma2-9BAttack success rate49.50%%Dynamic prompt rewriting#S5.T8; row Gemma2-9B; condition Dynamic prompt rewriting; column Attack success rate
Gemma2-27BAttack success rate98.75%%DPI baseline#S5.T8; row Gemma2-27B; condition DPI baseline; column Attack success rate
Gemma2-27BAttack success rate99.75%%Delimiter#S5.T8; row Gemma2-27B; condition Delimiter; column Attack success rate
Gemma2-27BAttack success rate68.00%%Paraphrase#S5.T8; row Gemma2-27B; condition Paraphrase; column Attack success rate
Gemma2-27BAttack success rate99.50%%Instructional#S5.T8; row Gemma2-27B; condition Instructional; column Attack success rate
Gemma2-27BAttack success rate67.50%%Dynamic prompt rewriting#S5.T8; row Gemma2-27B; condition Dynamic prompt rewriting; column Attack success rate
LLaMA3-8BAttack success rate33.75%%DPI baseline#S5.T8; row LLaMA3-8B; condition DPI baseline; column Attack success rate
LLaMA3-8BAttack success rate62.75%%Delimiter#S5.T8; row LLaMA3-8B; condition Delimiter; column Attack success rate
LLaMA3-8BAttack success rate28.50%%Paraphrase#S5.T8; row LLaMA3-8B; condition Paraphrase; column Attack success rate
LLaMA3-8BAttack success rate52.00%%Instructional#S5.T8; row LLaMA3-8B; condition Instructional; column Attack success rate
LLaMA3-8BAttack success rate23.15%%Dynamic prompt rewriting#S5.T8; row LLaMA3-8B; condition Dynamic prompt rewriting; column Attack success rate
LLaMA3-70BAttack success rate87.75%%DPI baseline#S5.T8; row LLaMA3-70B; condition DPI baseline; column Attack success rate
LLaMA3-70BAttack success rate88.25%%Delimiter#S5.T8; row LLaMA3-70B; condition Delimiter; column Attack success rate
LLaMA3-70BAttack success rate71.25%%Paraphrase#S5.T8; row LLaMA3-70B; condition Paraphrase; column Attack success rate
LLaMA3-70BAttack success rate87.25%%Instructional#S5.T8; row LLaMA3-70B; condition Instructional; column Attack success rate
LLaMA3-70BAttack success rate74.50%%Dynamic prompt rewriting#S5.T8; row LLaMA3-70B; condition Dynamic prompt rewriting; column Attack success rate
LLaMA3.1-8BAttack success rate64.25%%DPI baseline#S5.T8; row LLaMA3.1-8B; condition DPI baseline; column Attack success rate
LLaMA3.1-8BAttack success rate65.00%%Delimiter#S5.T8; row LLaMA3.1-8B; condition Delimiter; column Attack success rate
LLaMA3.1-8BAttack success rate42.50%%Paraphrase#S5.T8; row LLaMA3.1-8B; condition Paraphrase; column Attack success rate
LLaMA3.1-8BAttack success rate68.75%%Instructional#S5.T8; row LLaMA3.1-8B; condition Instructional; column Attack success rate
LLaMA3.1-8BAttack success rate26.70%%Dynamic prompt rewriting#S5.T8; row LLaMA3.1-8B; condition Dynamic prompt rewriting; column Attack success rate
LLaMA3.1-70BAttack success rate93.50%%DPI baseline#S5.T8; row LLaMA3.1-70B; condition DPI baseline; column Attack success rate
LLaMA3.1-70BAttack success rate92.75%%Delimiter#S5.T8; row LLaMA3.1-70B; condition Delimiter; column Attack success rate
LLaMA3.1-70BAttack success rate56.75%%Paraphrase#S5.T8; row LLaMA3.1-70B; condition Paraphrase; column Attack success rate
LLaMA3.1-70BAttack success rate90.50%%Instructional#S5.T8; row LLaMA3.1-70B; condition Instructional; column Attack success rate
LLaMA3.1-70BAttack success rate44.60%%Dynamic prompt rewriting#S5.T8; row LLaMA3.1-70B; condition Dynamic prompt rewriting; column Attack success rate
Mixtral-8x7BAttack success rate43.25%%DPI baseline#S5.T8; row Mixtral-8x7B; condition DPI baseline; column Attack success rate
Mixtral-8x7BAttack success rate43.00%%Delimiter#S5.T8; row Mixtral-8x7B; condition Delimiter; column Attack success rate
Mixtral-8x7BAttack success rate21.00%%Paraphrase#S5.T8; row Mixtral-8x7B; condition Paraphrase; column Attack success rate
Mixtral-8x7BAttack success rate34.00%%Instructional#S5.T8; row Mixtral-8x7B; condition Instructional; column Attack success rate
Mixtral-8x7BAttack success rate8.35%%Dynamic prompt rewriting#S5.T8; row Mixtral-8x7B; condition Dynamic prompt rewriting; column Attack success rate
Qwen2-7BAttack success rate73.50%%DPI baseline#S5.T8; row Qwen2-7B; condition DPI baseline; column Attack success rate
Qwen2-7BAttack success rate80.00%%Delimiter#S5.T8; row Qwen2-7B; condition Delimiter; column Attack success rate
Qwen2-7BAttack success rate46.25%%Paraphrase#S5.T8; row Qwen2-7B; condition Paraphrase; column Attack success rate
Qwen2-7BAttack success rate76.75%%Instructional#S5.T8; row Qwen2-7B; condition Instructional; column Attack success rate
Qwen2-7BAttack success rate33.30%%Dynamic prompt rewriting#S5.T8; row Qwen2-7B; condition Dynamic prompt rewriting; column Attack success rate
Qwen2-72BAttack success rate94.50%%DPI baseline#S5.T8; row Qwen2-72B; condition DPI baseline; column Attack success rate
Qwen2-72BAttack success rate95.00%%Delimiter#S5.T8; row Qwen2-72B; condition Delimiter; column Attack success rate
Qwen2-72BAttack success rate60.50%%Paraphrase#S5.T8; row Qwen2-72B; condition Paraphrase; column Attack success rate
Qwen2-72BAttack success rate95.50%%Instructional#S5.T8; row Qwen2-72B; condition Instructional; column Attack success rate
Qwen2-72BAttack success rate39.40%%Dynamic prompt rewriting#S5.T8; row Qwen2-72B; condition Dynamic prompt rewriting; column Attack success rate
Claude-3.5 SonnetAttack success rate87.75%%DPI baseline#S5.T8; row Claude-3.5 Sonnet; condition DPI baseline; column Attack success rate
Claude-3.5 SonnetAttack success rate79.00%%Delimiter#S5.T8; row Claude-3.5 Sonnet; condition Delimiter; column Attack success rate
Claude-3.5 SonnetAttack success rate65.25%%Paraphrase#S5.T8; row Claude-3.5 Sonnet; condition Paraphrase; column Attack success rate
Claude-3.5 SonnetAttack success rate70.25%%Instructional#S5.T8; row Claude-3.5 Sonnet; condition Instructional; column Attack success rate
Claude-3.5 SonnetAttack success rate35.80%%Dynamic prompt rewriting#S5.T8; row Claude-3.5 Sonnet; condition Dynamic prompt rewriting; column Attack success rate
GPT-3.5 TurboAttack success rate99.75%%DPI baseline#S5.T8; row GPT-3.5 Turbo; condition DPI baseline; column Attack success rate
GPT-3.5 TurboAttack success rate99.75%%Delimiter#S5.T8; row GPT-3.5 Turbo; condition Delimiter; column Attack success rate
GPT-3.5 TurboAttack success rate78.25%%Paraphrase#S5.T8; row GPT-3.5 Turbo; condition Paraphrase; column Attack success rate
GPT-3.5 TurboAttack success rate99.50%%Instructional#S5.T8; row GPT-3.5 Turbo; condition Instructional; column Attack success rate
GPT-3.5 TurboAttack success rate53.60%%Dynamic prompt rewriting#S5.T8; row GPT-3.5 Turbo; condition Dynamic prompt rewriting; column Attack success rate
GPT-4oAttack success rate55.50%%DPI baseline#S5.T8; row GPT-4o; condition DPI baseline; column Attack success rate
GPT-4oAttack success rate52.25%%Delimiter#S5.T8; row GPT-4o; condition Delimiter; column Attack success rate
GPT-4oAttack success rate62.50%%Paraphrase#S5.T8; row GPT-4o; condition Paraphrase; column Attack success rate
GPT-4oAttack success rate70.75%%Instructional#S5.T8; row GPT-4o; condition Instructional; column Attack success rate
GPT-4oAttack success rate60.20%%Dynamic prompt rewriting#S5.T8; row GPT-4o; condition Dynamic prompt rewriting; column Attack success rate
GPT-4o-miniAttack success rate95.75%%DPI baseline#S5.T8; row GPT-4o-mini; condition DPI baseline; column Attack success rate
GPT-4o-miniAttack success rate78.75%%Delimiter#S5.T8; row GPT-4o-mini; condition Delimiter; column Attack success rate
GPT-4o-miniAttack success rate76.00%%Paraphrase#S5.T8; row GPT-4o-mini; condition Paraphrase; column Attack success rate
GPT-4o-miniAttack success rate62.25%%Instructional#S5.T8; row GPT-4o-mini; condition Instructional; column Attack success rate
GPT-4o-miniAttack success rate61.25%%Dynamic prompt rewriting#S5.T8; row GPT-4o-mini; condition Dynamic prompt rewriting; column Attack success rate
AverageAttack success rate78.38%%DPI baseline#S5.T8; row Average; condition DPI baseline; column Attack success rate
AverageAttack success rate79.08%%Delimiter#S5.T8; row Average; condition Delimiter; column Attack success rate
AverageAttack success rate56.87%%Paraphrase#S5.T8; row Average; condition Paraphrase; column Attack success rate
AverageAttack success rate76.77%%Instructional#S5.T8; row Average; condition Instructional; column Attack success rate
AverageAttack success rate44.45%%Dynamic prompt rewriting#S5.T8; row Average; condition Dynamic prompt rewriting; column Attack success rate
Change versus table-specific baselineAttack success rate0.00%percentage pointsDPI baseline#S5.T8; row Δ; condition DPI baseline; column Attack success rate
Change versus table-specific baselineAttack success rate0.69%percentage pointsDelimiter#S5.T8; row Δ; condition Delimiter; column Attack success rate
Change versus table-specific baselineAttack success rate-21.52%percentage pointsParaphrase#S5.T8; row Δ; condition Paraphrase; column Attack success rate
Change versus table-specific baselineAttack success rate-1.62%percentage pointsInstructional#S5.T8; row Δ; condition Instructional; column Attack success rate
Change versus table-specific baselineAttack success rate-33.93%percentage pointsDynamic prompt rewriting#S5.T8; row Δ; condition Dynamic prompt rewriting; column Attack success rate
Gemma2-9BAttack success rate88.75%%DPI+IPI#A4.T14; row Gemma2-9B; condition DPI+IPI; column Attack success rate
Gemma2-9BRefusal rate3.75%%DPI+IPI#A4.T14; row Gemma2-9B; condition DPI+IPI; column Refusal rate
Gemma2-9BAttack success rate91.75%%DPI+MP#A4.T14; row Gemma2-9B; condition DPI+MP; column Attack success rate
Gemma2-9BRefusal rate1.75%%DPI+MP#A4.T14; row Gemma2-9B; condition DPI+MP; column Refusal rate
Gemma2-9BAttack success rate14.00%%IPI+MP#A4.T14; row Gemma2-9B; condition IPI+MP; column Attack success rate
Gemma2-9BRefusal rate26.75%%IPI+MP#A4.T14; row Gemma2-9B; condition IPI+MP; column Refusal rate
Gemma2-9BAttack success rate92.17%%DPI+IPI+MP#A4.T14; row Gemma2-9B; condition DPI+IPI+MP; column Attack success rate
Gemma2-9BRefusal rate1.33%%DPI+IPI+MP#A4.T14; row Gemma2-9B; condition DPI+IPI+MP; column Refusal rate
Gemma2-27BAttack success rate99.25%%DPI+IPI#A4.T14; row Gemma2-27B; condition DPI+IPI; column Attack success rate
Gemma2-27BRefusal rate0.00%%DPI+IPI#A4.T14; row Gemma2-27B; condition DPI+IPI; column Refusal rate
Gemma2-27BAttack success rate99.00%%DPI+MP#A4.T14; row Gemma2-27B; condition DPI+MP; column Attack success rate
Gemma2-27BRefusal rate1.00%%DPI+MP#A4.T14; row Gemma2-27B; condition DPI+MP; column Refusal rate
Gemma2-27BAttack success rate16.00%%IPI+MP#A4.T14; row Gemma2-27B; condition IPI+MP; column Attack success rate
Gemma2-27BRefusal rate3.25%%IPI+MP#A4.T14; row Gemma2-27B; condition IPI+MP; column Refusal rate
Gemma2-27BAttack success rate100.00%%DPI+IPI+MP#A4.T14; row Gemma2-27B; condition DPI+IPI+MP; column Attack success rate
Gemma2-27BRefusal rate0.50%%DPI+IPI+MP#A4.T14; row Gemma2-27B; condition DPI+IPI+MP; column Refusal rate
LLaMA3-8BAttack success rate55.75%%DPI+IPI#A4.T14; row LLaMA3-8B; condition DPI+IPI; column Attack success rate
LLaMA3-8BRefusal rate14.75%%DPI+IPI#A4.T14; row LLaMA3-8B; condition DPI+IPI; column Refusal rate
LLaMA3-8BAttack success rate37.75%%DPI+MP#A4.T14; row LLaMA3-8B; condition DPI+MP; column Attack success rate
LLaMA3-8BRefusal rate5.25%%DPI+MP#A4.T14; row LLaMA3-8B; condition DPI+MP; column Refusal rate
LLaMA3-8BAttack success rate4.25%%IPI+MP#A4.T14; row LLaMA3-8B; condition IPI+MP; column Attack success rate
LLaMA3-8BRefusal rate1.75%%IPI+MP#A4.T14; row LLaMA3-8B; condition IPI+MP; column Refusal rate
LLaMA3-8BAttack success rate40.75%%DPI+IPI+MP#A4.T14; row LLaMA3-8B; condition DPI+IPI+MP; column Attack success rate
LLaMA3-8BRefusal rate5.75%%DPI+IPI+MP#A4.T14; row LLaMA3-8B; condition DPI+IPI+MP; column Refusal rate
LLaMA3-70BAttack success rate86.50%%DPI+IPI#A4.T14; row LLaMA3-70B; condition DPI+IPI; column Attack success rate
LLaMA3-70BRefusal rate7.50%%DPI+IPI#A4.T14; row LLaMA3-70B; condition DPI+IPI; column Refusal rate
LLaMA3-70BAttack success rate83.75%%DPI+MP#A4.T14; row LLaMA3-70B; condition DPI+MP; column Attack success rate
LLaMA3-70BRefusal rate6.50%%DPI+MP#A4.T14; row LLaMA3-70B; condition DPI+MP; column Refusal rate
LLaMA3-70BAttack success rate27.50%%IPI+MP#A4.T14; row LLaMA3-70B; condition IPI+MP; column Attack success rate
LLaMA3-70BRefusal rate2.50%%IPI+MP#A4.T14; row LLaMA3-70B; condition IPI+MP; column Refusal rate
LLaMA3-70BAttack success rate85.50%%DPI+IPI+MP#A4.T14; row LLaMA3-70B; condition DPI+IPI+MP; column Attack success rate
LLaMA3-70BRefusal rate6.50%%DPI+IPI+MP#A4.T14; row LLaMA3-70B; condition DPI+IPI+MP; column Refusal rate
LLaMA3.1-8BAttack success rate68.25%%DPI+IPI#A4.T14; row LLaMA3.1-8B; condition DPI+IPI; column Attack success rate
LLaMA3.1-8BRefusal rate10.00%%DPI+IPI#A4.T14; row LLaMA3.1-8B; condition DPI+IPI; column Refusal rate
LLaMA3.1-8BAttack success rate67.50%%DPI+MP#A4.T14; row LLaMA3.1-8B; condition DPI+MP; column Attack success rate
LLaMA3.1-8BRefusal rate5.25%%DPI+MP#A4.T14; row LLaMA3.1-8B; condition DPI+MP; column Refusal rate
LLaMA3.1-8BAttack success rate6.25%%IPI+MP#A4.T14; row LLaMA3.1-8B; condition IPI+MP; column Attack success rate
LLaMA3.1-8BRefusal rate2.75%%IPI+MP#A4.T14; row LLaMA3.1-8B; condition IPI+MP; column Refusal rate
LLaMA3.1-8BAttack success rate73.50%%DPI+IPI+MP#A4.T14; row LLaMA3.1-8B; condition DPI+IPI+MP; column Attack success rate
LLaMA3.1-8BRefusal rate3.50%%DPI+IPI+MP#A4.T14; row LLaMA3.1-8B; condition DPI+IPI+MP; column Refusal rate
LLaMA3.1-70BAttack success rate88.00%%DPI+IPI#A4.T14; row LLaMA3.1-70B; condition DPI+IPI; column Attack success rate
LLaMA3.1-70BRefusal rate4.50%%DPI+IPI#A4.T14; row LLaMA3.1-70B; condition DPI+IPI; column Refusal rate
LLaMA3.1-70BAttack success rate95.75%%DPI+MP#A4.T14; row LLaMA3.1-70B; condition DPI+MP; column Attack success rate
LLaMA3.1-70BRefusal rate0.75%%DPI+MP#A4.T14; row LLaMA3.1-70B; condition DPI+MP; column Refusal rate
LLaMA3.1-70BAttack success rate10.25%%IPI+MP#A4.T14; row LLaMA3.1-70B; condition IPI+MP; column Attack success rate
LLaMA3.1-70BRefusal rate2.50%%IPI+MP#A4.T14; row LLaMA3.1-70B; condition IPI+MP; column Refusal rate
LLaMA3.1-70BAttack success rate94.50%%DPI+IPI+MP#A4.T14; row LLaMA3.1-70B; condition DPI+IPI+MP; column Attack success rate
LLaMA3.1-70BRefusal rate1.25%%DPI+IPI+MP#A4.T14; row LLaMA3.1-70B; condition DPI+IPI+MP; column Refusal rate
Mixtral-8x7BAttack success rate32.50%%DPI+IPI#A4.T14; row Mixtral-8x7B; condition DPI+IPI; column Attack success rate
Mixtral-8x7BRefusal rate11.25%%DPI+IPI#A4.T14; row Mixtral-8x7B; condition DPI+IPI; column Refusal rate
Mixtral-8x7BAttack success rate55.50%%DPI+MP#A4.T14; row Mixtral-8x7B; condition DPI+MP; column Attack success rate
Mixtral-8x7BRefusal rate7.75%%DPI+MP#A4.T14; row Mixtral-8x7B; condition DPI+MP; column Refusal rate
Mixtral-8x7BAttack success rate4.25%%IPI+MP#A4.T14; row Mixtral-8x7B; condition IPI+MP; column Attack success rate
Mixtral-8x7BRefusal rate5.50%%IPI+MP#A4.T14; row Mixtral-8x7B; condition IPI+MP; column Refusal rate
Mixtral-8x7BAttack success rate54.75%%DPI+IPI+MP#A4.T14; row Mixtral-8x7B; condition DPI+IPI+MP; column Attack success rate
Mixtral-8x7BRefusal rate6.75%%DPI+IPI+MP#A4.T14; row Mixtral-8x7B; condition DPI+IPI+MP; column Refusal rate
Qwen2-7BAttack success rate71.50%%DPI+IPI#A4.T14; row Qwen2-7B; condition DPI+IPI; column Attack success rate
Qwen2-7BRefusal rate6.50%%DPI+IPI#A4.T14; row Qwen2-7B; condition DPI+IPI; column Refusal rate
Qwen2-7BAttack success rate70.00%%DPI+MP#A4.T14; row Qwen2-7B; condition DPI+MP; column Attack success rate
Qwen2-7BRefusal rate1.50%%DPI+MP#A4.T14; row Qwen2-7B; condition DPI+MP; column Refusal rate
Qwen2-7BAttack success rate9.25%%IPI+MP#A4.T14; row Qwen2-7B; condition IPI+MP; column Attack success rate
Qwen2-7BRefusal rate7.00%%IPI+MP#A4.T14; row Qwen2-7B; condition IPI+MP; column Refusal rate
Qwen2-7BAttack success rate76.00%%DPI+IPI+MP#A4.T14; row Qwen2-7B; condition DPI+IPI+MP; column Attack success rate
Qwen2-7BRefusal rate2.50%%DPI+IPI+MP#A4.T14; row Qwen2-7B; condition DPI+IPI+MP; column Refusal rate
Qwen2-72BAttack success rate97.75%%DPI+IPI#A4.T14; row Qwen2-72B; condition DPI+IPI; column Attack success rate
Qwen2-72BRefusal rate1.75%%DPI+IPI#A4.T14; row Qwen2-72B; condition DPI+IPI; column Refusal rate
Qwen2-72BAttack success rate98.00%%DPI+MP#A4.T14; row Qwen2-72B; condition DPI+MP; column Attack success rate
Qwen2-72BRefusal rate3.00%%DPI+MP#A4.T14; row Qwen2-72B; condition DPI+MP; column Refusal rate
Qwen2-72BAttack success rate23.25%%IPI+MP#A4.T14; row Qwen2-72B; condition IPI+MP; column Attack success rate
Qwen2-72BRefusal rate1.75%%IPI+MP#A4.T14; row Qwen2-72B; condition IPI+MP; column Refusal rate
Qwen2-72BAttack success rate98.50%%DPI+IPI+MP#A4.T14; row Qwen2-72B; condition DPI+IPI+MP; column Attack success rate
Qwen2-72BRefusal rate0.75%%DPI+IPI+MP#A4.T14; row Qwen2-72B; condition DPI+IPI+MP; column Refusal rate
Claude-3.5 SonnetAttack success rate88.00%%DPI+IPI#A4.T14; row Claude-3.5 Sonnet; condition DPI+IPI; column Attack success rate
Claude-3.5 SonnetRefusal rate21.00%%DPI+IPI#A4.T14; row Claude-3.5 Sonnet; condition DPI+IPI; column Refusal rate
Claude-3.5 SonnetAttack success rate94.50%%DPI+MP#A4.T14; row Claude-3.5 Sonnet; condition DPI+MP; column Attack success rate
Claude-3.5 SonnetRefusal rate6.75%%DPI+MP#A4.T14; row Claude-3.5 Sonnet; condition DPI+MP; column Refusal rate
Claude-3.5 SonnetAttack success rate58.25%%IPI+MP#A4.T14; row Claude-3.5 Sonnet; condition IPI+MP; column Attack success rate
Claude-3.5 SonnetRefusal rate16.75%%IPI+MP#A4.T14; row Claude-3.5 Sonnet; condition IPI+MP; column Refusal rate
Claude-3.5 SonnetAttack success rate94.50%%DPI+IPI+MP#A4.T14; row Claude-3.5 Sonnet; condition DPI+IPI+MP; column Attack success rate
Claude-3.5 SonnetRefusal rate6.25%%DPI+IPI+MP#A4.T14; row Claude-3.5 Sonnet; condition DPI+IPI+MP; column Refusal rate
GPT-3.5 TurboAttack success rate99.50%%DPI+IPI#A4.T14; row GPT-3.5 Turbo; condition DPI+IPI; column Attack success rate
GPT-3.5 TurboRefusal rate0.00%%DPI+IPI#A4.T14; row GPT-3.5 Turbo; condition DPI+IPI; column Refusal rate
GPT-3.5 TurboAttack success rate100.00%%DPI+MP#A4.T14; row GPT-3.5 Turbo; condition DPI+MP; column Attack success rate
GPT-3.5 TurboRefusal rate0.25%%DPI+MP#A4.T14; row GPT-3.5 Turbo; condition DPI+MP; column Refusal rate
GPT-3.5 TurboAttack success rate41.25%%IPI+MP#A4.T14; row GPT-3.5 Turbo; condition IPI+MP; column Attack success rate
GPT-3.5 TurboRefusal rate12.75%%IPI+MP#A4.T14; row GPT-3.5 Turbo; condition IPI+MP; column Refusal rate
GPT-3.5 TurboAttack success rate99.75%%DPI+IPI+MP#A4.T14; row GPT-3.5 Turbo; condition DPI+IPI+MP; column Attack success rate
GPT-3.5 TurboRefusal rate0.00%%DPI+IPI+MP#A4.T14; row GPT-3.5 Turbo; condition DPI+IPI+MP; column Refusal rate
GPT-4oAttack success rate63.50%%DPI+IPI#A4.T14; row GPT-4o; condition DPI+IPI; column Attack success rate
GPT-4oRefusal rate19.25%%DPI+IPI#A4.T14; row GPT-4o; condition DPI+IPI; column Refusal rate
GPT-4oAttack success rate88.25%%DPI+MP#A4.T14; row GPT-4o; condition DPI+MP; column Attack success rate
GPT-4oRefusal rate7.50%%DPI+MP#A4.T14; row GPT-4o; condition DPI+MP; column Refusal rate
GPT-4oAttack success rate55.75%%IPI+MP#A4.T14; row GPT-4o; condition IPI+MP; column Attack success rate
GPT-4oRefusal rate0.75%%IPI+MP#A4.T14; row GPT-4o; condition IPI+MP; column Refusal rate
GPT-4oAttack success rate89.25%%DPI+IPI+MP#A4.T14; row GPT-4o; condition DPI+IPI+MP; column Attack success rate
GPT-4oRefusal rate5.50%%DPI+IPI+MP#A4.T14; row GPT-4o; condition DPI+IPI+MP; column Refusal rate
GPT-4o-miniAttack success rate88.50%%DPI+IPI#A4.T14; row GPT-4o-mini; condition DPI+IPI; column Attack success rate
GPT-4o-miniRefusal rate4.25%%DPI+IPI#A4.T14; row GPT-4o-mini; condition DPI+IPI; column Refusal rate
GPT-4o-miniAttack success rate97.50%%DPI+MP#A4.T14; row GPT-4o-mini; condition DPI+MP; column Attack success rate
GPT-4o-miniRefusal rate1.25%%DPI+MP#A4.T14; row GPT-4o-mini; condition DPI+MP; column Refusal rate
GPT-4o-miniAttack success rate38.50%%IPI+MP#A4.T14; row GPT-4o-mini; condition IPI+MP; column Attack success rate
GPT-4o-miniRefusal rate0.00%%IPI+MP#A4.T14; row GPT-4o-mini; condition IPI+MP; column Refusal rate
GPT-4o-miniAttack success rate96.75%%DPI+IPI+MP#A4.T14; row GPT-4o-mini; condition DPI+IPI+MP; column Attack success rate
GPT-4o-miniRefusal rate1.25%%DPI+IPI+MP#A4.T14; row GPT-4o-mini; condition DPI+IPI+MP; column Refusal rate
AverageAttack success rate79.06%%DPI+IPI#A4.T14; row Average; condition DPI+IPI; column Attack success rate
AverageRefusal rate8.04%%DPI+IPI#A4.T14; row Average; condition DPI+IPI; column Refusal rate
AverageAttack success rate83.02%%DPI+MP#A4.T14; row Average; condition DPI+MP; column Attack success rate
AverageRefusal rate3.73%%DPI+MP#A4.T14; row Average; condition DPI+MP; column Refusal rate
AverageAttack success rate23.75%%IPI+MP#A4.T14; row Average; condition IPI+MP; column Attack success rate
AverageRefusal rate6.46%%IPI+MP#A4.T14; row Average; condition IPI+MP; column Refusal rate
AverageAttack success rate84.30%%DPI+IPI+MP#A4.T14; row Average; condition DPI+IPI+MP; column Attack success rate
AverageRefusal rate3.22%%DPI+IPI+MP#A4.T14; row Average; condition DPI+IPI+MP; column Refusal rate
Change versus table-specific baselineAttack success rate6.37%percentage pointsDPI+IPI#A4.T14; row Δ; condition DPI+IPI; column Attack success rate
Change versus table-specific baselineRefusal rate1.50%percentage pointsDPI+IPI#A4.T14; row Δ; condition DPI+IPI; column Refusal rate
Change versus table-specific baselineAttack success rate10.33%percentage pointsDPI+MP#A4.T14; row Δ; condition DPI+MP; column Attack success rate
Change versus table-specific baselineRefusal rate-2.80%percentage pointsDPI+MP#A4.T14; row Δ; condition DPI+MP; column Refusal rate
Change versus table-specific baselineAttack success rate-3.80%percentage pointsIPI+MP#A4.T14; row Δ; condition IPI+MP; column Attack success rate
Change versus table-specific baselineRefusal rate-2.15%percentage pointsIPI+MP#A4.T14; row Δ; condition IPI+MP; column Refusal rate
Change versus table-specific baselineAttack success rate11.62%percentage pointsDPI+IPI+MP#A4.T14; row Δ; condition DPI+IPI+MP; column Attack success rate
Change versus table-specific baselineRefusal rate-3.32%percentage pointsDPI+IPI+MP#A4.T14; row Δ; condition DPI+IPI+MP; column Refusal rate
Gemma2-9BBenign performance21.00%%PoT backdoor#A4.T16; row Gemma2-9B; condition PoT backdoor; column Benign performance
Gemma2-9BPerformance under no attack10.75%%No attack#A4.T16; row Gemma2-9B; condition No attack; column Performance under no attack
Gemma2-27BBenign performance37.75%%PoT backdoor#A4.T16; row Gemma2-27B; condition PoT backdoor; column Benign performance
Gemma2-27BPerformance under no attack31.50%%No attack#A4.T16; row Gemma2-27B; condition No attack; column Performance under no attack
LLaMA3-8BBenign performance4.25%%PoT backdoor#A4.T16; row LLaMA3-8B; condition PoT backdoor; column Benign performance
LLaMA3-8BPerformance under no attack1.50%%No attack#A4.T16; row LLaMA3-8B; condition No attack; column Performance under no attack
LLaMA3-70BBenign performance59.00%%PoT backdoor#A4.T16; row LLaMA3-70B; condition PoT backdoor; column Benign performance
LLaMA3-70BPerformance under no attack66.50%%No attack#A4.T16; row LLaMA3-70B; condition No attack; column Performance under no attack
LLaMA3.1-8BBenign performance2.75%%PoT backdoor#A4.T16; row LLaMA3.1-8B; condition PoT backdoor; column Benign performance
LLaMA3.1-8BPerformance under no attack0.75%%No attack#A4.T16; row LLaMA3.1-8B; condition No attack; column Performance under no attack
LLaMA3.1-70BBenign performance32.00%%PoT backdoor#A4.T16; row LLaMA3.1-70B; condition PoT backdoor; column Benign performance
LLaMA3.1-70BPerformance under no attack21.25%%No attack#A4.T16; row LLaMA3.1-70B; condition No attack; column Performance under no attack
Mixtral-8x7BBenign performance1.00%%PoT backdoor#A4.T16; row Mixtral-8x7B; condition PoT backdoor; column Benign performance
Mixtral-8x7BPerformance under no attack0.00%%No attack#A4.T16; row Mixtral-8x7B; condition No attack; column Performance under no attack
Qwen2-7BBenign performance8.50%%PoT backdoor#A4.T16; row Qwen2-7B; condition PoT backdoor; column Benign performance
Qwen2-7BPerformance under no attack9.75%%No attack#A4.T16; row Qwen2-7B; condition No attack; column Performance under no attack
Qwen2-72BBenign performance2.25%%PoT backdoor#A4.T16; row Qwen2-72B; condition PoT backdoor; column Benign performance
Qwen2-72BPerformance under no attack4.00%%No attack#A4.T16; row Qwen2-72B; condition No attack; column Performance under no attack
Claude-3.5 SonnetBenign performance90.75%%PoT backdoor#A4.T16; row Claude-3.5 Sonnet; condition PoT backdoor; column Benign performance
Claude-3.5 SonnetPerformance under no attack100.00%%No attack#A4.T16; row Claude-3.5 Sonnet; condition No attack; column Performance under no attack
GPT-3.5 TurboBenign performance17.25%%PoT backdoor#A4.T16; row GPT-3.5 Turbo; condition PoT backdoor; column Benign performance
GPT-3.5 TurboPerformance under no attack8.00%%No attack#A4.T16; row GPT-3.5 Turbo; condition No attack; column Performance under no attack
GPT-4oBenign performance77.75%%PoT backdoor#A4.T16; row GPT-4o; condition PoT backdoor; column Benign performance
GPT-4oPerformance under no attack79.00%%No attack#A4.T16; row GPT-4o; condition No attack; column Performance under no attack
GPT-4o-miniBenign performance64.50%%PoT backdoor#A4.T16; row GPT-4o-mini; condition PoT backdoor; column Benign performance
GPT-4o-miniPerformance under no attack50.00%%No attack#A4.T16; row GPT-4o-mini; condition No attack; column Performance under no attack
AverageBenign performance32.21%%PoT backdoor#A4.T16; row Average; condition PoT backdoor; column Benign performance
AveragePerformance under no attack29.46%%No attack#A4.T16; row Average; condition No attack; column Performance under no attack
Gemma2-9BPerformance under no attack10.75%%No attack#A4.T19; row Gemma2-9B; condition No attack; column Performance under no attack
Gemma2-9BPerformance under no attack8.00%%Delimiters#A4.T19; row Gemma2-9B; condition Delimiters; column Performance under no attack
Gemma2-9BChange from no-defense PNA-2.75%percentage pointsDelimiters#A4.T19; row Gemma2-9B; condition Delimiters; column Change from no-defense PNA
Gemma2-9BPerformance under no attack9.00%%Paraphrasing#A4.T19; row Gemma2-9B; condition Paraphrasing; column Performance under no attack
Gemma2-9BChange from no-defense PNA-1.75%percentage pointsParaphrasing#A4.T19; row Gemma2-9B; condition Paraphrasing; column Change from no-defense PNA
Gemma2-9BPerformance under no attack8.00%%Instructional#A4.T19; row Gemma2-9B; condition Instructional; column Performance under no attack
Gemma2-9BChange from no-defense PNA-2.75%percentage pointsInstructional#A4.T19; row Gemma2-9B; condition Instructional; column Change from no-defense PNA
Gemma2-9BPerformance under no attack14.00%%Sandwich#A4.T19; row Gemma2-9B; condition Sandwich; column Performance under no attack
Gemma2-9BChange from no-defense PNA3.25%percentage pointsSandwich#A4.T19; row Gemma2-9B; condition Sandwich; column Change from no-defense PNA
Gemma2-9BPerformance under no attack10.20%%DPR#A4.T19; row Gemma2-9B; condition DPR; column Performance under no attack
Gemma2-9BChange from no-defense PNA-0.55%percentage pointsDPR#A4.T19; row Gemma2-9B; condition DPR; column Change from no-defense PNA
Gemma2-27BPerformance under no attack31.50%%No attack#A4.T19; row Gemma2-27B; condition No attack; column Performance under no attack
Gemma2-27BPerformance under no attack14.75%%Delimiters#A4.T19; row Gemma2-27B; condition Delimiters; column Performance under no attack
Gemma2-27BChange from no-defense PNA-16.75%percentage pointsDelimiters#A4.T19; row Gemma2-27B; condition Delimiters; column Change from no-defense PNA
Gemma2-27BPerformance under no attack24.75%%Paraphrasing#A4.T19; row Gemma2-27B; condition Paraphrasing; column Performance under no attack
Gemma2-27BChange from no-defense PNA-6.75%percentage pointsParaphrasing#A4.T19; row Gemma2-27B; condition Paraphrasing; column Change from no-defense PNA
Gemma2-27BPerformance under no attack23.00%%Instructional#A4.T19; row Gemma2-27B; condition Instructional; column Performance under no attack
Gemma2-27BChange from no-defense PNA-8.50%percentage pointsInstructional#A4.T19; row Gemma2-27B; condition Instructional; column Change from no-defense PNA
Gemma2-27BPerformance under no attack27.00%%Sandwich#A4.T19; row Gemma2-27B; condition Sandwich; column Performance under no attack
Gemma2-27BChange from no-defense PNA-4.50%percentage pointsSandwich#A4.T19; row Gemma2-27B; condition Sandwich; column Change from no-defense PNA
Gemma2-27BPerformance under no attack30.50%%DPR#A4.T19; row Gemma2-27B; condition DPR; column Performance under no attack
Gemma2-27BChange from no-defense PNA-1.00%percentage pointsDPR#A4.T19; row Gemma2-27B; condition DPR; column Change from no-defense PNA
LLaMA3-8BPerformance under no attack1.50%%No attack#A4.T19; row LLaMA3-8B; condition No attack; column Performance under no attack
LLaMA3-8BPerformance under no attack1.75%%Delimiters#A4.T19; row LLaMA3-8B; condition Delimiters; column Performance under no attack
LLaMA3-8BChange from no-defense PNA0.25%percentage pointsDelimiters#A4.T19; row LLaMA3-8B; condition Delimiters; column Change from no-defense PNA
LLaMA3-8BPerformance under no attack8.25%%Paraphrasing#A4.T19; row LLaMA3-8B; condition Paraphrasing; column Performance under no attack
LLaMA3-8BChange from no-defense PNA6.75%percentage pointsParaphrasing#A4.T19; row LLaMA3-8B; condition Paraphrasing; column Change from no-defense PNA
LLaMA3-8BPerformance under no attack5.00%%Instructional#A4.T19; row LLaMA3-8B; condition Instructional; column Performance under no attack
LLaMA3-8BChange from no-defense PNA3.50%percentage pointsInstructional#A4.T19; row LLaMA3-8B; condition Instructional; column Change from no-defense PNA
LLaMA3-8BPerformance under no attack8.50%%Sandwich#A4.T19; row LLaMA3-8B; condition Sandwich; column Performance under no attack
LLaMA3-8BChange from no-defense PNA7.00%percentage pointsSandwich#A4.T19; row LLaMA3-8B; condition Sandwich; column Change from no-defense PNA
LLaMA3-8BPerformance under no attack7.50%%DPR#A4.T19; row LLaMA3-8B; condition DPR; column Performance under no attack
LLaMA3-8BChange from no-defense PNA6.00%percentage pointsDPR#A4.T19; row LLaMA3-8B; condition DPR; column Change from no-defense PNA
LLaMA3-70BPerformance under no attack66.50%%No attack#A4.T19; row LLaMA3-70B; condition No attack; column Performance under no attack
LLaMA3-70BPerformance under no attack58.50%%Delimiters#A4.T19; row LLaMA3-70B; condition Delimiters; column Performance under no attack
LLaMA3-70BChange from no-defense PNA-8.00%percentage pointsDelimiters#A4.T19; row LLaMA3-70B; condition Delimiters; column Change from no-defense PNA
LLaMA3-70BPerformance under no attack68.25%%Paraphrasing#A4.T19; row LLaMA3-70B; condition Paraphrasing; column Performance under no attack
LLaMA3-70BChange from no-defense PNA1.75%percentage pointsParaphrasing#A4.T19; row LLaMA3-70B; condition Paraphrasing; column Change from no-defense PNA
LLaMA3-70BPerformance under no attack70.00%%Instructional#A4.T19; row LLaMA3-70B; condition Instructional; column Performance under no attack
LLaMA3-70BChange from no-defense PNA3.50%percentage pointsInstructional#A4.T19; row LLaMA3-70B; condition Instructional; column Change from no-defense PNA
LLaMA3-70BPerformance under no attack68.00%%Sandwich#A4.T19; row LLaMA3-70B; condition Sandwich; column Performance under no attack
LLaMA3-70BChange from no-defense PNA1.50%percentage pointsSandwich#A4.T19; row LLaMA3-70B; condition Sandwich; column Change from no-defense PNA
LLaMA3-70BPerformance under no attack64.50%%DPR#A4.T19; row LLaMA3-70B; condition DPR; column Performance under no attack
LLaMA3-70BChange from no-defense PNA-2.00%percentage pointsDPR#A4.T19; row LLaMA3-70B; condition DPR; column Change from no-defense PNA
LLaMA3.1-8BPerformance under no attack0.75%%No attack#A4.T19; row LLaMA3.1-8B; condition No attack; column Performance under no attack
LLaMA3.1-8BPerformance under no attack0.00%%Delimiters#A4.T19; row LLaMA3.1-8B; condition Delimiters; column Performance under no attack
LLaMA3.1-8BChange from no-defense PNA-0.75%percentage pointsDelimiters#A4.T19; row LLaMA3.1-8B; condition Delimiters; column Change from no-defense PNA
LLaMA3.1-8BPerformance under no attack1.00%%Paraphrasing#A4.T19; row LLaMA3.1-8B; condition Paraphrasing; column Performance under no attack
LLaMA3.1-8BChange from no-defense PNA0.25%percentage pointsParaphrasing#A4.T19; row LLaMA3.1-8B; condition Paraphrasing; column Change from no-defense PNA
LLaMA3.1-8BPerformance under no attack0.75%%Instructional#A4.T19; row LLaMA3.1-8B; condition Instructional; column Performance under no attack
LLaMA3.1-8BChange from no-defense PNA0.00%percentage pointsInstructional#A4.T19; row LLaMA3.1-8B; condition Instructional; column Change from no-defense PNA
LLaMA3.1-8BPerformance under no attack0.25%%Sandwich#A4.T19; row LLaMA3.1-8B; condition Sandwich; column Performance under no attack
LLaMA3.1-8BChange from no-defense PNA-0.50%percentage pointsSandwich#A4.T19; row LLaMA3.1-8B; condition Sandwich; column Change from no-defense PNA
LLaMA3.1-8BPerformance under no attack0.50%%DPR#A4.T19; row LLaMA3.1-8B; condition DPR; column Performance under no attack
LLaMA3.1-8BChange from no-defense PNA-0.25%percentage pointsDPR#A4.T19; row LLaMA3.1-8B; condition DPR; column Change from no-defense PNA
LLaMA3.1-70BPerformance under no attack21.25%%No attack#A4.T19; row LLaMA3.1-70B; condition No attack; column Performance under no attack
LLaMA3.1-70BPerformance under no attack9.50%%Delimiters#A4.T19; row LLaMA3.1-70B; condition Delimiters; column Performance under no attack
LLaMA3.1-70BChange from no-defense PNA-11.75%percentage pointsDelimiters#A4.T19; row LLaMA3.1-70B; condition Delimiters; column Change from no-defense PNA
LLaMA3.1-70BPerformance under no attack18.75%%Paraphrasing#A4.T19; row LLaMA3.1-70B; condition Paraphrasing; column Performance under no attack
LLaMA3.1-70BChange from no-defense PNA-2.50%percentage pointsParaphrasing#A4.T19; row LLaMA3.1-70B; condition Paraphrasing; column Change from no-defense PNA
LLaMA3.1-70BPerformance under no attack14.50%%Instructional#A4.T19; row LLaMA3.1-70B; condition Instructional; column Performance under no attack
LLaMA3.1-70BChange from no-defense PNA-6.75%percentage pointsInstructional#A4.T19; row LLaMA3.1-70B; condition Instructional; column Change from no-defense PNA
LLaMA3.1-70BPerformance under no attack16.50%%Sandwich#A4.T19; row LLaMA3.1-70B; condition Sandwich; column Performance under no attack
LLaMA3.1-70BChange from no-defense PNA-4.75%percentage pointsSandwich#A4.T19; row LLaMA3.1-70B; condition Sandwich; column Change from no-defense PNA
LLaMA3.1-70BPerformance under no attack13.80%%DPR#A4.T19; row LLaMA3.1-70B; condition DPR; column Performance under no attack
LLaMA3.1-70BChange from no-defense PNA-7.45%percentage pointsDPR#A4.T19; row LLaMA3.1-70B; condition DPR; column Change from no-defense PNA
Mixtral-8x7BPerformance under no attack0.00%%No attack#A4.T19; row Mixtral-8x7B; condition No attack; column Performance under no attack
Mixtral-8x7BPerformance under no attack0.25%%Delimiters#A4.T19; row Mixtral-8x7B; condition Delimiters; column Performance under no attack
Mixtral-8x7BChange from no-defense PNA0.25%percentage pointsDelimiters#A4.T19; row Mixtral-8x7B; condition Delimiters; column Change from no-defense PNA
Mixtral-8x7BPerformance under no attack0.00%%Paraphrasing#A4.T19; row Mixtral-8x7B; condition Paraphrasing; column Performance under no attack
Mixtral-8x7BChange from no-defense PNA0.00%percentage pointsParaphrasing#A4.T19; row Mixtral-8x7B; condition Paraphrasing; column Change from no-defense PNA
Mixtral-8x7BPerformance under no attack0.00%%Instructional#A4.T19; row Mixtral-8x7B; condition Instructional; column Performance under no attack
Mixtral-8x7BChange from no-defense PNA0.00%percentage pointsInstructional#A4.T19; row Mixtral-8x7B; condition Instructional; column Change from no-defense PNA
Mixtral-8x7BPerformance under no attack0.25%%Sandwich#A4.T19; row Mixtral-8x7B; condition Sandwich; column Performance under no attack
Mixtral-8x7BChange from no-defense PNA0.25%percentage pointsSandwich#A4.T19; row Mixtral-8x7B; condition Sandwich; column Change from no-defense PNA
Mixtral-8x7BPerformance under no attack0.30%%DPR#A4.T19; row Mixtral-8x7B; condition DPR; column Performance under no attack
Mixtral-8x7BChange from no-defense PNA0.30%percentage pointsDPR#A4.T19; row Mixtral-8x7B; condition DPR; column Change from no-defense PNA
Qwen2-7BPerformance under no attack9.75%%No attack#A4.T19; row Qwen2-7B; condition No attack; column Performance under no attack
Qwen2-7BPerformance under no attack1.75%%Delimiters#A4.T19; row Qwen2-7B; condition Delimiters; column Performance under no attack
Qwen2-7BChange from no-defense PNA-8.00%percentage pointsDelimiters#A4.T19; row Qwen2-7B; condition Delimiters; column Change from no-defense PNA
Qwen2-7BPerformance under no attack7.00%%Paraphrasing#A4.T19; row Qwen2-7B; condition Paraphrasing; column Performance under no attack
Qwen2-7BChange from no-defense PNA-2.75%percentage pointsParaphrasing#A4.T19; row Qwen2-7B; condition Paraphrasing; column Change from no-defense PNA
Qwen2-7BPerformance under no attack5.75%%Instructional#A4.T19; row Qwen2-7B; condition Instructional; column Performance under no attack
Qwen2-7BChange from no-defense PNA-4.00%percentage pointsInstructional#A4.T19; row Qwen2-7B; condition Instructional; column Change from no-defense PNA
Qwen2-7BPerformance under no attack5.25%%Sandwich#A4.T19; row Qwen2-7B; condition Sandwich; column Performance under no attack
Qwen2-7BChange from no-defense PNA-4.50%percentage pointsSandwich#A4.T19; row Qwen2-7B; condition Sandwich; column Change from no-defense PNA
Qwen2-7BPerformance under no attack6.80%%DPR#A4.T19; row Qwen2-7B; condition DPR; column Performance under no attack
Qwen2-7BChange from no-defense PNA-2.95%percentage pointsDPR#A4.T19; row Qwen2-7B; condition DPR; column Change from no-defense PNA
Qwen2-72BPerformance under no attack4.00%%No attack#A4.T19; row Qwen2-72B; condition No attack; column Performance under no attack
Qwen2-72BPerformance under no attack0.25%%Delimiters#A4.T19; row Qwen2-72B; condition Delimiters; column Performance under no attack
Qwen2-72BChange from no-defense PNA-3.75%percentage pointsDelimiters#A4.T19; row Qwen2-72B; condition Delimiters; column Change from no-defense PNA
Qwen2-72BPerformance under no attack0.00%%Paraphrasing#A4.T19; row Qwen2-72B; condition Paraphrasing; column Performance under no attack
Qwen2-72BChange from no-defense PNA-4.00%percentage pointsParaphrasing#A4.T19; row Qwen2-72B; condition Paraphrasing; column Change from no-defense PNA
Qwen2-72BPerformance under no attack0.50%%Instructional#A4.T19; row Qwen2-72B; condition Instructional; column Performance under no attack
Qwen2-72BChange from no-defense PNA-3.50%percentage pointsInstructional#A4.T19; row Qwen2-72B; condition Instructional; column Change from no-defense PNA
Qwen2-72BPerformance under no attack0.50%%Sandwich#A4.T19; row Qwen2-72B; condition Sandwich; column Performance under no attack
Qwen2-72BChange from no-defense PNA-3.50%percentage pointsSandwich#A4.T19; row Qwen2-72B; condition Sandwich; column Change from no-defense PNA
Qwen2-72BPerformance under no attack0.00%%DPR#A4.T19; row Qwen2-72B; condition DPR; column Performance under no attack
Qwen2-72BChange from no-defense PNA-4.00%percentage pointsDPR#A4.T19; row Qwen2-72B; condition DPR; column Change from no-defense PNA
Claude-3.5 SonnetPerformance under no attack100.00%%No attack#A4.T19; row Claude-3.5 Sonnet; condition No attack; column Performance under no attack
Claude-3.5 SonnetPerformance under no attack90.00%%Delimiters#A4.T19; row Claude-3.5 Sonnet; condition Delimiters; column Performance under no attack
Claude-3.5 SonnetChange from no-defense PNA-10.00%percentage pointsDelimiters#A4.T19; row Claude-3.5 Sonnet; condition Delimiters; column Change from no-defense PNA
Claude-3.5 SonnetPerformance under no attack99.75%%Paraphrasing#A4.T19; row Claude-3.5 Sonnet; condition Paraphrasing; column Performance under no attack
Claude-3.5 SonnetChange from no-defense PNA-0.25%percentage pointsParaphrasing#A4.T19; row Claude-3.5 Sonnet; condition Paraphrasing; column Change from no-defense PNA
Claude-3.5 SonnetPerformance under no attack90.00%%Instructional#A4.T19; row Claude-3.5 Sonnet; condition Instructional; column Performance under no attack
Claude-3.5 SonnetChange from no-defense PNA-10.00%percentage pointsInstructional#A4.T19; row Claude-3.5 Sonnet; condition Instructional; column Change from no-defense PNA
Claude-3.5 SonnetPerformance under no attack100.00%%Sandwich#A4.T19; row Claude-3.5 Sonnet; condition Sandwich; column Performance under no attack
Claude-3.5 SonnetChange from no-defense PNA0.00%percentage pointsSandwich#A4.T19; row Claude-3.5 Sonnet; condition Sandwich; column Change from no-defense PNA
Claude-3.5 SonnetPerformance under no attack95.80%%DPR#A4.T19; row Claude-3.5 Sonnet; condition DPR; column Performance under no attack
Claude-3.5 SonnetChange from no-defense PNA-4.20%percentage pointsDPR#A4.T19; row Claude-3.5 Sonnet; condition DPR; column Change from no-defense PNA
GPT-3.5 TurboPerformance under no attack8.00%%No attack#A4.T19; row GPT-3.5 Turbo; condition No attack; column Performance under no attack
GPT-3.5 TurboPerformance under no attack0.75%%Delimiters#A4.T19; row GPT-3.5 Turbo; condition Delimiters; column Performance under no attack
GPT-3.5 TurboChange from no-defense PNA-7.25%percentage pointsDelimiters#A4.T19; row GPT-3.5 Turbo; condition Delimiters; column Change from no-defense PNA
GPT-3.5 TurboPerformance under no attack6.75%%Paraphrasing#A4.T19; row GPT-3.5 Turbo; condition Paraphrasing; column Performance under no attack
GPT-3.5 TurboChange from no-defense PNA-1.25%percentage pointsParaphrasing#A4.T19; row GPT-3.5 Turbo; condition Paraphrasing; column Change from no-defense PNA
GPT-3.5 TurboPerformance under no attack13.50%%Instructional#A4.T19; row GPT-3.5 Turbo; condition Instructional; column Performance under no attack
GPT-3.5 TurboChange from no-defense PNA5.50%percentage pointsInstructional#A4.T19; row GPT-3.5 Turbo; condition Instructional; column Change from no-defense PNA
GPT-3.5 TurboPerformance under no attack10.25%%Sandwich#A4.T19; row GPT-3.5 Turbo; condition Sandwich; column Performance under no attack
GPT-3.5 TurboChange from no-defense PNA2.25%percentage pointsSandwich#A4.T19; row GPT-3.5 Turbo; condition Sandwich; column Change from no-defense PNA
GPT-3.5 TurboPerformance under no attack9.50%%DPR#A4.T19; row GPT-3.5 Turbo; condition DPR; column Performance under no attack
GPT-3.5 TurboChange from no-defense PNA1.50%percentage pointsDPR#A4.T19; row GPT-3.5 Turbo; condition DPR; column Change from no-defense PNA
GPT-4oPerformance under no attack79.00%%No attack#A4.T19; row GPT-4o; condition No attack; column Performance under no attack
GPT-4oPerformance under no attack70.25%%Delimiters#A4.T19; row GPT-4o; condition Delimiters; column Performance under no attack
GPT-4oChange from no-defense PNA-8.75%percentage pointsDelimiters#A4.T19; row GPT-4o; condition Delimiters; column Change from no-defense PNA
GPT-4oPerformance under no attack80.00%%Paraphrasing#A4.T19; row GPT-4o; condition Paraphrasing; column Performance under no attack
GPT-4oChange from no-defense PNA1.00%percentage pointsParaphrasing#A4.T19; row GPT-4o; condition Paraphrasing; column Change from no-defense PNA
GPT-4oPerformance under no attack72.75%%Instructional#A4.T19; row GPT-4o; condition Instructional; column Performance under no attack
GPT-4oChange from no-defense PNA-6.25%percentage pointsInstructional#A4.T19; row GPT-4o; condition Instructional; column Change from no-defense PNA
GPT-4oPerformance under no attack79.50%%Sandwich#A4.T19; row GPT-4o; condition Sandwich; column Performance under no attack
GPT-4oChange from no-defense PNA0.50%percentage pointsSandwich#A4.T19; row GPT-4o; condition Sandwich; column Change from no-defense PNA
GPT-4oPerformance under no attack73.80%%DPR#A4.T19; row GPT-4o; condition DPR; column Performance under no attack
GPT-4oChange from no-defense PNA-5.20%percentage pointsDPR#A4.T19; row GPT-4o; condition DPR; column Change from no-defense PNA
GPT-4o-miniPerformance under no attack50.00%%No attack#A4.T19; row GPT-4o-mini; condition No attack; column Performance under no attack
GPT-4o-miniPerformance under no attack37.00%%Delimiters#A4.T19; row GPT-4o-mini; condition Delimiters; column Performance under no attack
GPT-4o-miniChange from no-defense PNA-13.00%percentage pointsDelimiters#A4.T19; row GPT-4o-mini; condition Delimiters; column Change from no-defense PNA
GPT-4o-miniPerformance under no attack36.50%%Paraphrasing#A4.T19; row GPT-4o-mini; condition Paraphrasing; column Performance under no attack
GPT-4o-miniChange from no-defense PNA-13.50%percentage pointsParaphrasing#A4.T19; row GPT-4o-mini; condition Paraphrasing; column Change from no-defense PNA
GPT-4o-miniPerformance under no attack42.50%%Instructional#A4.T19; row GPT-4o-mini; condition Instructional; column Performance under no attack
GPT-4o-miniChange from no-defense PNA-7.50%percentage pointsInstructional#A4.T19; row GPT-4o-mini; condition Instructional; column Change from no-defense PNA
GPT-4o-miniPerformance under no attack37.75%%Sandwich#A4.T19; row GPT-4o-mini; condition Sandwich; column Performance under no attack
GPT-4o-miniChange from no-defense PNA-12.25%percentage pointsSandwich#A4.T19; row GPT-4o-mini; condition Sandwich; column Change from no-defense PNA
GPT-4o-miniPerformance under no attack29.70%%DPR#A4.T19; row GPT-4o-mini; condition DPR; column Performance under no attack
GPT-4o-miniChange from no-defense PNA-20.30%percentage pointsDPR#A4.T19; row GPT-4o-mini; condition DPR; column Change from no-defense PNA
AveragePerformance under no attack29.46%%No attack#A4.T19; row Average; condition No attack; column Performance under no attack
AveragePerformance under no attack22.52%%Delimiters#A4.T19; row Average; condition Delimiters; column Performance under no attack
AverageChange from no-defense PNA-6.94%percentage pointsDelimiters#A4.T19; row Average; condition Delimiters; column Change from no-defense PNA
AveragePerformance under no attack27.69%%Paraphrasing#A4.T19; row Average; condition Paraphrasing; column Performance under no attack
AverageChange from no-defense PNA-1.77%percentage pointsParaphrasing#A4.T19; row Average; condition Paraphrasing; column Change from no-defense PNA
AveragePerformance under no attack26.63%%Instructional#A4.T19; row Average; condition Instructional; column Performance under no attack
AverageChange from no-defense PNA-2.83%percentage pointsInstructional#A4.T19; row Average; condition Instructional; column Change from no-defense PNA
AveragePerformance under no attack28.29%%Sandwich#A4.T19; row Average; condition Sandwich; column Performance under no attack
AverageChange from no-defense PNA-1.17%percentage pointsSandwich#A4.T19; row Average; condition Sandwich; column Change from no-defense PNA
AveragePerformance under no attack26.38%%DPR#A4.T19; row Average; condition DPR; column Performance under no attack
AverageChange from no-defense PNA-3.08%percentage pointsDPR#A4.T19; row Average; condition DPR; column Change from no-defense PNA
Gemma2-9BAttack success rate39.75%%PoT baseline#A4.T20; row Gemma2-9B; condition PoT baseline; column Attack success rate
Gemma2-9BPerformance under no attack10.75%%PoT baseline#A4.T20; row Gemma2-9B; condition PoT baseline; column Performance under no attack
Gemma2-9BAttack success rate67.25%%Shuffle defense#A4.T20; row Gemma2-9B; condition Shuffle defense; column Attack success rate
Gemma2-9BPerformance under no attack22.25%%Shuffle defense#A4.T20; row Gemma2-9B; condition Shuffle defense; column Performance under no attack
Gemma2-9BAttack success rate24.50%%Paraphrase defense#A4.T20; row Gemma2-9B; condition Paraphrase defense; column Attack success rate
Gemma2-9BPerformance under no attack21.75%%Paraphrase defense#A4.T20; row Gemma2-9B; condition Paraphrase defense; column Performance under no attack
Gemma2-27BAttack success rate54.50%%PoT baseline#A4.T20; row Gemma2-27B; condition PoT baseline; column Attack success rate
Gemma2-27BPerformance under no attack31.50%%PoT baseline#A4.T20; row Gemma2-27B; condition PoT baseline; column Performance under no attack
Gemma2-27BAttack success rate59.50%%Shuffle defense#A4.T20; row Gemma2-27B; condition Shuffle defense; column Attack success rate
Gemma2-27BPerformance under no attack40.75%%Shuffle defense#A4.T20; row Gemma2-27B; condition Shuffle defense; column Performance under no attack
Gemma2-27BAttack success rate23.25%%Paraphrase defense#A4.T20; row Gemma2-27B; condition Paraphrase defense; column Attack success rate
Gemma2-27BPerformance under no attack32.25%%Paraphrase defense#A4.T20; row Gemma2-27B; condition Paraphrase defense; column Performance under no attack
LLaMA3-8BAttack success rate21.50%%PoT baseline#A4.T20; row LLaMA3-8B; condition PoT baseline; column Attack success rate
LLaMA3-8BPerformance under no attack1.50%%PoT baseline#A4.T20; row LLaMA3-8B; condition PoT baseline; column Performance under no attack
LLaMA3-8BAttack success rate2.25%%Shuffle defense#A4.T20; row LLaMA3-8B; condition Shuffle defense; column Attack success rate
LLaMA3-8BPerformance under no attack3.50%%Shuffle defense#A4.T20; row LLaMA3-8B; condition Shuffle defense; column Performance under no attack
LLaMA3-8BAttack success rate5.00%%Paraphrase defense#A4.T20; row LLaMA3-8B; condition Paraphrase defense; column Attack success rate
LLaMA3-8BPerformance under no attack6.00%%Paraphrase defense#A4.T20; row LLaMA3-8B; condition Paraphrase defense; column Performance under no attack
LLaMA3-70BAttack success rate57.00%%PoT baseline#A4.T20; row LLaMA3-70B; condition PoT baseline; column Attack success rate
LLaMA3-70BPerformance under no attack66.50%%PoT baseline#A4.T20; row LLaMA3-70B; condition PoT baseline; column Performance under no attack
LLaMA3-70BAttack success rate63.75%%Shuffle defense#A4.T20; row LLaMA3-70B; condition Shuffle defense; column Attack success rate
LLaMA3-70BPerformance under no attack54.50%%Shuffle defense#A4.T20; row LLaMA3-70B; condition Shuffle defense; column Performance under no attack
LLaMA3-70BAttack success rate44.75%%Paraphrase defense#A4.T20; row LLaMA3-70B; condition Paraphrase defense; column Attack success rate
LLaMA3-70BPerformance under no attack52.75%%Paraphrase defense#A4.T20; row LLaMA3-70B; condition Paraphrase defense; column Performance under no attack
LLaMA3.1-8BAttack success rate19.00%%PoT baseline#A4.T20; row LLaMA3.1-8B; condition PoT baseline; column Attack success rate
LLaMA3.1-8BPerformance under no attack0.75%%PoT baseline#A4.T20; row LLaMA3.1-8B; condition PoT baseline; column Performance under no attack
LLaMA3.1-8BAttack success rate17.25%%Shuffle defense#A4.T20; row LLaMA3.1-8B; condition Shuffle defense; column Attack success rate
LLaMA3.1-8BPerformance under no attack2.75%%Shuffle defense#A4.T20; row LLaMA3.1-8B; condition Shuffle defense; column Performance under no attack
LLaMA3.1-8BAttack success rate17.50%%Paraphrase defense#A4.T20; row LLaMA3.1-8B; condition Paraphrase defense; column Attack success rate
LLaMA3.1-8BPerformance under no attack2.50%%Paraphrase defense#A4.T20; row LLaMA3.1-8B; condition Paraphrase defense; column Performance under no attack
LLaMA3.1-70BAttack success rate59.75%%PoT baseline#A4.T20; row LLaMA3.1-70B; condition PoT baseline; column Attack success rate
LLaMA3.1-70BPerformance under no attack21.25%%PoT baseline#A4.T20; row LLaMA3.1-70B; condition PoT baseline; column Performance under no attack
LLaMA3.1-70BAttack success rate69.00%%Shuffle defense#A4.T20; row LLaMA3.1-70B; condition Shuffle defense; column Attack success rate
LLaMA3.1-70BPerformance under no attack43.00%%Shuffle defense#A4.T20; row LLaMA3.1-70B; condition Shuffle defense; column Performance under no attack
LLaMA3.1-70BAttack success rate42.00%%Paraphrase defense#A4.T20; row LLaMA3.1-70B; condition Paraphrase defense; column Attack success rate
LLaMA3.1-70BPerformance under no attack30.00%%Paraphrase defense#A4.T20; row LLaMA3.1-70B; condition Paraphrase defense; column Performance under no attack
Mixtral-8x7BAttack success rate4.75%%PoT baseline#A4.T20; row Mixtral-8x7B; condition PoT baseline; column Attack success rate
Mixtral-8x7BPerformance under no attack0.00%%PoT baseline#A4.T20; row Mixtral-8x7B; condition PoT baseline; column Performance under no attack
Mixtral-8x7BAttack success rate12.25%%Shuffle defense#A4.T20; row Mixtral-8x7B; condition Shuffle defense; column Attack success rate
Mixtral-8x7BPerformance under no attack0.25%%Shuffle defense#A4.T20; row Mixtral-8x7B; condition Shuffle defense; column Performance under no attack
Mixtral-8x7BAttack success rate4.50%%Paraphrase defense#A4.T20; row Mixtral-8x7B; condition Paraphrase defense; column Attack success rate
Mixtral-8x7BPerformance under no attack0.50%%Paraphrase defense#A4.T20; row Mixtral-8x7B; condition Paraphrase defense; column Performance under no attack
Qwen2-7BAttack success rate12.25%%PoT baseline#A4.T20; row Qwen2-7B; condition PoT baseline; column Attack success rate
Qwen2-7BPerformance under no attack9.75%%PoT baseline#A4.T20; row Qwen2-7B; condition PoT baseline; column Performance under no attack
Qwen2-7BAttack success rate14.50%%Shuffle defense#A4.T20; row Qwen2-7B; condition Shuffle defense; column Attack success rate
Qwen2-7BPerformance under no attack13.00%%Shuffle defense#A4.T20; row Qwen2-7B; condition Shuffle defense; column Performance under no attack
Qwen2-7BAttack success rate11.00%%Paraphrase defense#A4.T20; row Qwen2-7B; condition Paraphrase defense; column Attack success rate
Qwen2-7BPerformance under no attack10.25%%Paraphrase defense#A4.T20; row Qwen2-7B; condition Paraphrase defense; column Performance under no attack
Qwen2-72BAttack success rate57.75%%PoT baseline#A4.T20; row Qwen2-72B; condition PoT baseline; column Attack success rate
Qwen2-72BPerformance under no attack4.00%%PoT baseline#A4.T20; row Qwen2-72B; condition PoT baseline; column Performance under no attack
Qwen2-72BAttack success rate22.75%%Shuffle defense#A4.T20; row Qwen2-72B; condition Shuffle defense; column Attack success rate
Qwen2-72BPerformance under no attack10.75%%Shuffle defense#A4.T20; row Qwen2-72B; condition Shuffle defense; column Performance under no attack
Qwen2-72BAttack success rate37.75%%Paraphrase defense#A4.T20; row Qwen2-72B; condition Paraphrase defense; column Attack success rate
Qwen2-72BPerformance under no attack18.00%%Paraphrase defense#A4.T20; row Qwen2-72B; condition Paraphrase defense; column Performance under no attack
Claude-3.5 SonnetAttack success rate17.50%%PoT baseline#A4.T20; row Claude-3.5 Sonnet; condition PoT baseline; column Attack success rate
Claude-3.5 SonnetPerformance under no attack100.00%%PoT baseline#A4.T20; row Claude-3.5 Sonnet; condition PoT baseline; column Performance under no attack
Claude-3.5 SonnetAttack success rate93.50%%Shuffle defense#A4.T20; row Claude-3.5 Sonnet; condition Shuffle defense; column Attack success rate
Claude-3.5 SonnetPerformance under no attack81.50%%Shuffle defense#A4.T20; row Claude-3.5 Sonnet; condition Shuffle defense; column Performance under no attack
Claude-3.5 SonnetAttack success rate13.75%%Paraphrase defense#A4.T20; row Claude-3.5 Sonnet; condition Paraphrase defense; column Attack success rate
Claude-3.5 SonnetPerformance under no attack82.75%%Paraphrase defense#A4.T20; row Claude-3.5 Sonnet; condition Paraphrase defense; column Performance under no attack
GPT-3.5 TurboAttack success rate8.25%%PoT baseline#A4.T20; row GPT-3.5 Turbo; condition PoT baseline; column Attack success rate
GPT-3.5 TurboPerformance under no attack8.00%%PoT baseline#A4.T20; row GPT-3.5 Turbo; condition PoT baseline; column Performance under no attack
GPT-3.5 TurboAttack success rate16.50%%Shuffle defense#A4.T20; row GPT-3.5 Turbo; condition Shuffle defense; column Attack success rate
GPT-3.5 TurboPerformance under no attack16.75%%Shuffle defense#A4.T20; row GPT-3.5 Turbo; condition Shuffle defense; column Performance under no attack
GPT-3.5 TurboAttack success rate6.25%%Paraphrase defense#A4.T20; row GPT-3.5 Turbo; condition Paraphrase defense; column Attack success rate
GPT-3.5 TurboPerformance under no attack23.50%%Paraphrase defense#A4.T20; row GPT-3.5 Turbo; condition Paraphrase defense; column Performance under no attack
GPT-4oAttack success rate100.00%%PoT baseline#A4.T20; row GPT-4o; condition PoT baseline; column Attack success rate
GPT-4oPerformance under no attack79.00%%PoT baseline#A4.T20; row GPT-4o; condition PoT baseline; column Performance under no attack
GPT-4oAttack success rate98.50%%Shuffle defense#A4.T20; row GPT-4o; condition Shuffle defense; column Attack success rate
GPT-4oPerformance under no attack78.50%%Shuffle defense#A4.T20; row GPT-4o; condition Shuffle defense; column Performance under no attack
GPT-4oAttack success rate84.75%%Paraphrase defense#A4.T20; row GPT-4o; condition Paraphrase defense; column Attack success rate
GPT-4oPerformance under no attack88.00%%Paraphrase defense#A4.T20; row GPT-4o; condition Paraphrase defense; column Performance under no attack
GPT-4o-miniAttack success rate95.50%%PoT baseline#A4.T20; row GPT-4o-mini; condition PoT baseline; column Attack success rate
GPT-4o-miniPerformance under no attack50.00%%PoT baseline#A4.T20; row GPT-4o-mini; condition PoT baseline; column Performance under no attack
GPT-4o-miniAttack success rate39.75%%Shuffle defense#A4.T20; row GPT-4o-mini; condition Shuffle defense; column Attack success rate
GPT-4o-miniPerformance under no attack63.75%%Shuffle defense#A4.T20; row GPT-4o-mini; condition Shuffle defense; column Performance under no attack
GPT-4o-miniAttack success rate62.75%%Paraphrase defense#A4.T20; row GPT-4o-mini; condition Paraphrase defense; column Attack success rate
GPT-4o-miniPerformance under no attack79.00%%Paraphrase defense#A4.T20; row GPT-4o-mini; condition Paraphrase defense; column Performance under no attack
AverageAttack success rate42.12%%PoT baseline#A4.T20; row Average; condition PoT baseline; column Attack success rate
AveragePerformance under no attack29.46%%PoT baseline#A4.T20; row Average; condition PoT baseline; column Performance under no attack
AverageAttack success rate44.37%%Shuffle defense#A4.T20; row Average; condition Shuffle defense; column Attack success rate
AveragePerformance under no attack33.17%%Shuffle defense#A4.T20; row Average; condition Shuffle defense; column Performance under no attack
AverageAttack success rate29.06%%Paraphrase defense#A4.T20; row Average; condition Paraphrase defense; column Attack success rate
AveragePerformance under no attack34.40%%Paraphrase defense#A4.T20; row Average; condition Paraphrase defense; column Performance under no attack
Gemma2-9BFalse negative rate0.658ratioMemory attack detection#A4.T21; row Gemma2-9B; condition Memory attack detection; column False negative rate
Gemma2-9BFalse positive rate0.204ratioMemory attack detection#A4.T21; row Gemma2-9B; condition Memory attack detection; column False positive rate
Gemma2-27BFalse negative rate0.655ratioMemory attack detection#A4.T21; row Gemma2-27B; condition Memory attack detection; column False negative rate
Gemma2-27BFalse positive rate0.201ratioMemory attack detection#A4.T21; row Gemma2-27B; condition Memory attack detection; column False positive rate
LLaMA3-8BFalse negative rate0.654ratioMemory attack detection#A4.T21; row LLaMA3-8B; condition Memory attack detection; column False negative rate
LLaMA3-8BFalse positive rate0.204ratioMemory attack detection#A4.T21; row LLaMA3-8B; condition Memory attack detection; column False positive rate
LLaMA3-70BFalse negative rate0.661ratioMemory attack detection#A4.T21; row LLaMA3-70B; condition Memory attack detection; column False negative rate
LLaMA3-70BFalse positive rate0.202ratioMemory attack detection#A4.T21; row LLaMA3-70B; condition Memory attack detection; column False positive rate
LLaMA3.1-8BFalse negative rate0.656ratioMemory attack detection#A4.T21; row LLaMA3.1-8B; condition Memory attack detection; column False negative rate
LLaMA3.1-8BFalse positive rate0.200ratioMemory attack detection#A4.T21; row LLaMA3.1-8B; condition Memory attack detection; column False positive rate
LLaMA3.1-70BFalse negative rate0.659ratioMemory attack detection#A4.T21; row LLaMA3.1-70B; condition Memory attack detection; column False negative rate
LLaMA3.1-70BFalse positive rate0.197ratioMemory attack detection#A4.T21; row LLaMA3.1-70B; condition Memory attack detection; column False positive rate
Mixtral-8x7BFalse negative rate0.665ratioMemory attack detection#A4.T21; row Mixtral-8x7B; condition Memory attack detection; column False negative rate
Mixtral-8x7BFalse positive rate0.203ratioMemory attack detection#A4.T21; row Mixtral-8x7B; condition Memory attack detection; column False positive rate
Qwen2-7BFalse negative rate0.657ratioMemory attack detection#A4.T21; row Qwen2-7B; condition Memory attack detection; column False negative rate
Qwen2-7BFalse positive rate0.193ratioMemory attack detection#A4.T21; row Qwen2-7B; condition Memory attack detection; column False positive rate
Qwen2-72BFalse negative rate0.671ratioMemory attack detection#A4.T21; row Qwen2-72B; condition Memory attack detection; column False negative rate
Qwen2-72BFalse positive rate0.198ratioMemory attack detection#A4.T21; row Qwen2-72B; condition Memory attack detection; column False positive rate
Claude-3.5 SonnetFalse negative rate0.663ratioMemory attack detection#A4.T21; row Claude-3.5 Sonnet; condition Memory attack detection; column False negative rate
Claude-3.5 SonnetFalse positive rate0.199ratioMemory attack detection#A4.T21; row Claude-3.5 Sonnet; condition Memory attack detection; column False positive rate
GPT-3.5 TurboFalse negative rate0.661ratioMemory attack detection#A4.T21; row GPT-3.5 Turbo; condition Memory attack detection; column False negative rate
GPT-3.5 TurboFalse positive rate0.198ratioMemory attack detection#A4.T21; row GPT-3.5 Turbo; condition Memory attack detection; column False positive rate
GPT-4oFalse negative rate0.664ratioMemory attack detection#A4.T21; row GPT-4o; condition Memory attack detection; column False negative rate
GPT-4oFalse positive rate0.203ratioMemory attack detection#A4.T21; row GPT-4o; condition Memory attack detection; column False positive rate
GPT-4o-miniFalse negative rate0.657ratioMemory attack detection#A4.T21; row GPT-4o-mini; condition Memory attack detection; column False negative rate
GPT-4o-miniFalse positive rate0.200ratioMemory attack detection#A4.T21; row GPT-4o-mini; condition Memory attack detection; column False positive rate
AverageFalse negative rate0.660ratioMemory attack detection#A4.T21; row Average; condition Memory attack detection; column False negative rate
AverageFalse positive rate0.200ratioMemory attack detection#A4.T21; row Average; condition Memory attack detection; column False positive rate
Aggregate across evaluated backbonesAttack success rate78.38%%Combined Attack / DPI#A4.T17; row Aggregate across evaluated backbones; condition Combined Attack / DPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate5.35%%Combined Attack / DPI#A4.T17; row Aggregate across evaluated backbones; condition Combined Attack / DPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate27.98%%Combined Attack / IPI#A4.T17; row Aggregate across evaluated backbones; condition Combined Attack / IPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate10.27%%Combined Attack / IPI#A4.T17; row Aggregate across evaluated backbones; condition Combined Attack / IPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate7.65%%Combined Attack / Memory poisoning#A4.T17; row Aggregate across evaluated backbones; condition Combined Attack / Memory poisoning; column Attack success rate
Aggregate across evaluated backbonesRefusal rate4.75%%Combined Attack / Memory poisoning#A4.T17; row Aggregate across evaluated backbones; condition Combined Attack / Memory poisoning; column Refusal rate
Aggregate across evaluated backbonesAttack success rate38.01%%Combined Attack / Average#A4.T17; row Aggregate across evaluated backbones; condition Combined Attack / Average; column Attack success rate
Aggregate across evaluated backbonesRefusal rate6.79%%Combined Attack / Average#A4.T17; row Aggregate across evaluated backbones; condition Combined Attack / Average; column Refusal rate
Aggregate across evaluated backbonesAttack success rate73.85%%Context Ignoring / DPI#A4.T17; row Aggregate across evaluated backbones; condition Context Ignoring / DPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate6.33%%Context Ignoring / DPI#A4.T17; row Aggregate across evaluated backbones; condition Context Ignoring / DPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate27.29%%Context Ignoring / IPI#A4.T17; row Aggregate across evaluated backbones; condition Context Ignoring / IPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate9.50%%Context Ignoring / IPI#A4.T17; row Aggregate across evaluated backbones; condition Context Ignoring / IPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate8.52%%Context Ignoring / Memory poisoning#A4.T17; row Aggregate across evaluated backbones; condition Context Ignoring / Memory poisoning; column Attack success rate
Aggregate across evaluated backbonesRefusal rate4.58%%Context Ignoring / Memory poisoning#A4.T17; row Aggregate across evaluated backbones; condition Context Ignoring / Memory poisoning; column Refusal rate
Aggregate across evaluated backbonesAttack success rate36.55%%Context Ignoring / Average#A4.T17; row Aggregate across evaluated backbones; condition Context Ignoring / Average; column Attack success rate
Aggregate across evaluated backbonesRefusal rate6.80%%Context Ignoring / Average#A4.T17; row Aggregate across evaluated backbones; condition Context Ignoring / Average; column Refusal rate
Aggregate across evaluated backbonesAttack success rate68.73%%Escape Characters / DPI#A4.T17; row Aggregate across evaluated backbones; condition Escape Characters / DPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate7.21%%Escape Characters / DPI#A4.T17; row Aggregate across evaluated backbones; condition Escape Characters / DPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate27.81%%Escape Characters / IPI#A4.T17; row Aggregate across evaluated backbones; condition Escape Characters / IPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate7.62%%Escape Characters / IPI#A4.T17; row Aggregate across evaluated backbones; condition Escape Characters / IPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate7.71%%Escape Characters / Memory poisoning#A4.T17; row Aggregate across evaluated backbones; condition Escape Characters / Memory poisoning; column Attack success rate
Aggregate across evaluated backbonesRefusal rate4.38%%Escape Characters / Memory poisoning#A4.T17; row Aggregate across evaluated backbones; condition Escape Characters / Memory poisoning; column Refusal rate
Aggregate across evaluated backbonesAttack success rate34.75%%Escape Characters / Average#A4.T17; row Aggregate across evaluated backbones; condition Escape Characters / Average; column Attack success rate
Aggregate across evaluated backbonesRefusal rate6.40%%Escape Characters / Average#A4.T17; row Aggregate across evaluated backbones; condition Escape Characters / Average; column Refusal rate
Aggregate across evaluated backbonesAttack success rate71.94%%Fake Completion / DPI#A4.T17; row Aggregate across evaluated backbones; condition Fake Completion / DPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate5.63%%Fake Completion / DPI#A4.T17; row Aggregate across evaluated backbones; condition Fake Completion / DPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate26.62%%Fake Completion / IPI#A4.T17; row Aggregate across evaluated backbones; condition Fake Completion / IPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate8.19%%Fake Completion / IPI#A4.T17; row Aggregate across evaluated backbones; condition Fake Completion / IPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate8.00%%Fake Completion / Memory poisoning#A4.T17; row Aggregate across evaluated backbones; condition Fake Completion / Memory poisoning; column Attack success rate
Aggregate across evaluated backbonesRefusal rate4.87%%Fake Completion / Memory poisoning#A4.T17; row Aggregate across evaluated backbones; condition Fake Completion / Memory poisoning; column Refusal rate
Aggregate across evaluated backbonesAttack success rate35.52%%Fake Completion / Average#A4.T17; row Aggregate across evaluated backbones; condition Fake Completion / Average; column Attack success rate
Aggregate across evaluated backbonesRefusal rate6.23%%Fake Completion / Average#A4.T17; row Aggregate across evaluated backbones; condition Fake Completion / Average; column Refusal rate
Aggregate across evaluated backbonesAttack success rate70.52%%Naive / DPI#A4.T17; row Aggregate across evaluated backbones; condition Naive / DPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate8.15%%Naive / DPI#A4.T17; row Aggregate across evaluated backbones; condition Naive / DPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate28.04%%Naive / IPI#A4.T17; row Aggregate across evaluated backbones; condition Naive / IPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate7.46%%Naive / IPI#A4.T17; row Aggregate across evaluated backbones; condition Naive / IPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate7.73%%Naive / Memory poisoning#A4.T17; row Aggregate across evaluated backbones; condition Naive / Memory poisoning; column Attack success rate
Aggregate across evaluated backbonesRefusal rate4.56%%Naive / Memory poisoning#A4.T17; row Aggregate across evaluated backbones; condition Naive / Memory poisoning; column Refusal rate
Aggregate across evaluated backbonesAttack success rate35.43%%Naive / Average#A4.T17; row Aggregate across evaluated backbones; condition Naive / Average; column Attack success rate
Aggregate across evaluated backbonesRefusal rate6.72%%Naive / Average#A4.T17; row Aggregate across evaluated backbones; condition Naive / Average; column Refusal rate
Aggregate across evaluated backbonesAttack success rate72.68%%Average / DPI#A4.T17; row Aggregate across evaluated backbones; condition Average / DPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate6.53%%Average / DPI#A4.T17; row Aggregate across evaluated backbones; condition Average / DPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate27.55%%Average / IPI#A4.T17; row Aggregate across evaluated backbones; condition Average / IPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate8.61%%Average / IPI#A4.T17; row Aggregate across evaluated backbones; condition Average / IPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate7.92%%Average / Memory poisoning#A4.T17; row Aggregate across evaluated backbones; condition Average / Memory poisoning; column Attack success rate
Aggregate across evaluated backbonesRefusal rate4.63%%Average / Memory poisoning#A4.T17; row Aggregate across evaluated backbones; condition Average / Memory poisoning; column Refusal rate
Aggregate across evaluated backbonesAttack success rate36.05%%Average / Average#A4.T17; row Aggregate across evaluated backbones; condition Average / Average; column Attack success rate
Aggregate across evaluated backbonesRefusal rate6.59%%Average / Average#A4.T17; row Aggregate across evaluated backbones; condition Average / Average; column Refusal rate
Aggregate across evaluated backbonesAttack success rate74.59%%No / DPI#A4.T18; row Aggregate across evaluated backbones; condition No / DPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate3.51%%No / DPI#A4.T18; row Aggregate across evaluated backbones; condition No / DPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate32.67%%No / IPI#A4.T18; row Aggregate across evaluated backbones; condition No / IPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate6.52%%No / IPI#A4.T18; row Aggregate across evaluated backbones; condition No / IPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate9.68%%No / Memory poisoning#A4.T18; row Aggregate across evaluated backbones; condition No / Memory poisoning; column Attack success rate
Aggregate across evaluated backbonesRefusal rate4.57%%No / Memory poisoning#A4.T18; row Aggregate across evaluated backbones; condition No / Memory poisoning; column Refusal rate
Aggregate across evaluated backbonesAttack success rate38.98%%No / Average#A4.T18; row Aggregate across evaluated backbones; condition No / Average; column Attack success rate
Aggregate across evaluated backbonesRefusal rate4.87%%No / Average#A4.T18; row Aggregate across evaluated backbones; condition No / Average; column Refusal rate
Aggregate across evaluated backbonesAttack success rate70.78%%Yes / DPI#A4.T18; row Aggregate across evaluated backbones; condition Yes / DPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate9.56%%Yes / DPI#A4.T18; row Aggregate across evaluated backbones; condition Yes / DPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate22.42%%Yes / IPI#A4.T18; row Aggregate across evaluated backbones; condition Yes / IPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate10.69%%Yes / IPI#A4.T18; row Aggregate across evaluated backbones; condition Yes / IPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate6.16%%Yes / Memory poisoning#A4.T18; row Aggregate across evaluated backbones; condition Yes / Memory poisoning; column Attack success rate
Aggregate across evaluated backbonesRefusal rate4.68%%Yes / Memory poisoning#A4.T18; row Aggregate across evaluated backbones; condition Yes / Memory poisoning; column Refusal rate
Aggregate across evaluated backbonesAttack success rate33.12%%Yes / Average#A4.T18; row Aggregate across evaluated backbones; condition Yes / Average; column Attack success rate
Aggregate across evaluated backbonesRefusal rate8.31%%Yes / Average#A4.T18; row Aggregate across evaluated backbones; condition Yes / Average; column Refusal rate
Aggregate across evaluated backbonesAttack success rate72.68%%Average / DPI#A4.T18; row Aggregate across evaluated backbones; condition Average / DPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate6.53%%Average / DPI#A4.T18; row Aggregate across evaluated backbones; condition Average / DPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate27.55%%Average / IPI#A4.T18; row Aggregate across evaluated backbones; condition Average / IPI; column Attack success rate
Aggregate across evaluated backbonesRefusal rate8.61%%Average / IPI#A4.T18; row Aggregate across evaluated backbones; condition Average / IPI; column Refusal rate
Aggregate across evaluated backbonesAttack success rate7.92%%Average / Memory poisoning#A4.T18; row Aggregate across evaluated backbones; condition Average / Memory poisoning; column Attack success rate
Aggregate across evaluated backbonesRefusal rate4.63%%Average / Memory poisoning#A4.T18; row Aggregate across evaluated backbones; condition Average / Memory poisoning; column Refusal rate
Aggregate across evaluated backbonesAttack success rate36.05%%Average / Average#A4.T18; row Aggregate across evaluated backbones; condition Average / Average; column Attack success rate
Aggregate across evaluated backbonesRefusal rate6.59%%Average / Average#A4.T18; row Aggregate across evaluated backbones; condition Average / Average; column Refusal rate
阅读范围与复核

Selected original HTML sections and table context; not a complete paper/appendix audit or independent reproduction.

模型辅助整理,独立证据复核待完成。

核对原始来源 ↗