Prompt Injection Defense on VPI-Bench
96.5ASR (Amazon)None
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| NoneAgent Framework=Browser-use Agent (Gemini-2.0-pro)2026.04 | 96.5 | 84.2 | 84.2 | |
| System PromptAgent Framework=Browser-use Agent (Gemini-2.0-pro)2026.04 | 92.98 | 85.96 | 85.96 | |
| System PromptAgent Framework=Claude Agent (Sonnet-3.7)2026.04 | 42.2 | 37.8 | 5.6 | |
| NoneAgent Framework=Claude Agent (Sonnet-3.7)2026.04 | 31.7 | 36.7 | 16.7 | |
| Guard-gpt-4oAgent Framework=Browser-use Agent (Gemini-2.0-pro)2026.04 | 22.8 | 15.8 | 21.1 | |
| Guard-gpt-4oAgent Framework=Claude Agent (Sonnet-3.7)2026.04 | 10.6 | 12.2 | 4.4 | |
| WebAgentGuard-8BAgent Framework=Claude Agent (Sonnet-3.7)2026.04 | 1.7 | 0 | 0 | |
| WebAgentGuard-4BAgent Framework=Claude Agent (Sonnet-3.7)2026.04 | 0 | 0.6 | 0 | |
| WebAgentGuard-8BAgent Framework=Browser-use Agent (Gemini-2.0-pro)2026.04 | 0 | 0 | 0 | |
| WebAgentGuard-4BAgent Framework=Browser-use Agent (Gemini-2.0-pro)2026.04 | 0 | 1.8 | 0 |