List your company

Jailbreak Detection on Average of six attacks

0Avg Success RateGradSafe
-3.7621.624772.38Dec 3, 2025Dec 22, 2025Jan 11, 2026Jan 31, 2026Feb 20, 2026Mar 12, 2026Apr 1, 2026

Evaluation results

MethodLinks
GradSafeTarget LLM=Vicuna-7B2025.120—
Self-ExTarget LLM=Vicuna-13B2025.120—
GradSafeTarget LLM=Vicuna-13B2025.120—
Self-ExTarget LLM=Vicuna-7B2025.122—
OAPITarget LLM=Vicuna-7B2025.125—
OAPITarget LLM=Vicuna-13B2025.125—
OAPITarget LLM=Llama2-7B2025.125—
OAPITarget LLM=Mistral-7B2025.126—
OAPITarget LLM=Llama3-8B2025.126—
GradSafeTarget LLM=Llama3-8B2025.127—
SelfGraderQ=10, Latency (Sec.)=0.74, Memory Overhead (MB)=2765.652026.047.4518.5
SelfDefendDefense Protocol=Intent, Latency (Sec.)=2.14, Memory Overhead (MB)=13194.112026.048.9529.96
GradSafeTarget LLM=Mistral-7B2025.1212—
SelfDefendDefense Protocol=Direct, Latency (Sec.)=0.84, Memory Overhead (MB)=13183.912026.0413.6836.41
Self-ExTarget LLM=Llama3-8B2025.1219—
PPLTarget LLM=Mistral-7B2025.1222—
Prompt GuardLatency (Sec.)=15.11, Memory Overhead (MB)=1064.732026.0423.5146.53
PPLTarget LLM=Vicuna-7B2025.1227—
Self-ExTarget LLM=Llama2-7B2025.1228—
PPLTarget LLM=Vicuna-13B2025.1230—
PPLTarget LLM=Llama2-7B2025.1231—
Llama GuardDeployment Stage=Post, Latency (Sec.)=0.56, Memory Overhead (MB)=13753.872026.0435.7777.73
Llama GuardDeployment Stage=Pre, Latency (Sec.)=0.49, Memory Overhead (MB)=13649.822026.0436.0574.86
Perplexity FilterLatency (Sec.)=0.17, Memory Overhead (MB)=13214.012026.0440.6184
PPLTarget LLM=Llama3-8B2025.1250—
LlamaGTarget LLM=Llama2-7B2025.1252—
Self-ExTarget LLM=Mistral-7B2025.1254—
Vicuna-13B-v1.5Defense=No Defense, Latency (Sec.)=3.202026.0454.94—
LlamaGTarget LLM=Llama3-8B2025.1259—
GradSafeTarget LLM=Llama2-7B2025.1264—
LlamaGTarget LLM=Vicuna-13B2025.1273—
LlamaGTarget LLM=Vicuna-7B2025.1274—
IMAGTarget LLM=Llama2-7B2025.1274—
LlamaGTarget LLM=Mistral-7B2025.1275—
IMAGTarget LLM=Vicuna-7B2025.1281—
IMAGTarget LLM=Llama3-8B2025.1285—
IMAGTarget LLM=Vicuna-13B2025.1293—
IMAGTarget LLM=Mistral-7B2025.1294—