List your company

Prompt Injection Defense on PI (CSQA) random topology

50ASR @1No Defense
10.79220.97131.1541.329Jan 21, 2026

Evaluation results

MethodLinks
No DefenseBackbone=Qwen3-235B-A22B2026.015070.779.351.728.316.7
AgentXposed-KickBackbone=Qwen3-235B-A22B2026.0147.770.778552516.7
AgentXposed-GuideBackbone=Qwen3-235B-A22B2026.0147.370.77758.326.718.3
InspectorBackbone=Qwen3-235B-A22B2026.0146.565.376.351.73521.7
AgentSafeBackbone=Qwen3-235B-A22B2026.0144.665.475.8553521.7
AgentXposed-GuideBackbone=GPT-4o-mini2026.0143.75155.356.75046.7
AgentSafeBackbone=GPT-4o-mini2026.0137.548.955.663.348.340
AgentXposed-KickBackbone=GPT-4o-mini2026.013745.34961.753.351.7
No DefenseBackbone=GPT-4o-mini2026.013650.259.365.350.341.7
G-safeguardBackbone=GPT-4o-mini2026.0131.730.731.766.77068.3
InspectorBackbone=GPT-4o-mini2026.0127.227.126.969.273.173.1
ChallengerBackbone=GPT-4o-mini2026.0125.34145.963.35551.7
INFA-GUARDBackbone=GPT-4o-mini2026.0124.323.323.37576.776.7
ChallengerBackbone=Qwen3-235B-A22B2026.0121.156.372.858.328.318.3
G-safeguardBackbone=Qwen3-235B-A22B2026.012124.326.383.376.778.3
INFA-GUARDBackbone=Qwen3-235B-A22B2026.0112.312.713.49086.786.7