Agentic Prompt Injection on ClawTrojan (positive)
95.5ASRGPT-5.4 (no defense)
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| GPT-5.4 (no defense)Latency=43.0s, Base Agent=GPT-5.42026.05 | 95.5 | 89.1 | 92.2 | 97.7 | 95.4 | 93.3 | 96.5 | |
| ClawKeeperLatency=18.8s, Base Agent=GPT-5.42026.05 | 94.3 | 86.1 | 90.3 | 94.3 | 94 | 90.3 | 98.3 | |
| StruQLatency=27.9s, Base Agent=GPT-5.42026.05 | 93.8 | 85.5 | 89.1 | 98.3 | 93.1 | 88.2 | 96.2 | |
| MELONLatency=30.1s, Base Agent=GPT-5.42026.05 | 92.6 | 83.8 | 87.5 | 93.2 | 90.7 | 90.8 | 95.5 | |
| PromptShield-1BLatency=18.0s, Base Agent=GPT-5.4, Model Scale=1B2026.05 | 91.2 | 81.4 | 85.9 | 89.2 | 88.4 | 89.9 | 95.8 | |
| GLM-5.1 (no defense)Latency=62.3s, Base Agent=GLM-5.12026.05 | 90.1 | 76.7 | 82.7 | 89.2 | 89.4 | 88.2 | 92.7 | |
| PromptShield-8BLatency=18.5s, Base Agent=GPT-5.4, Model Scale=8B2026.05 | 88.7 | 77.9 | 82.7 | 84.1 | 88.9 | 84 | 95.5 | |
| DS-V4-Flash (no defense)Latency=35.2s, Base Agent=DeepSeek-V4-Flash2026.05 | 88 | 74.3 | 82.4 | 86.9 | 88 | 85.3 | 90.9 | |
| CaMeLLatency=16.4s, Base Agent=GPT-5.42026.05 | 74.3 | 52.5 | 62.3 | 63.6 | 68.5 | 68.9 | 90.2 | |
| DASGuardLatency=31.0s, Base Agent=GPT-5.42026.05 | 15.8 | 5.9 | 10.1 | 29 | 11.1 | 14.7 | 12.2 |