Tool-use LLM Agent Task Completion on AgentDojo Workspace suite
87.5Benign Success RateRepeat User Prompt
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Repeat User PromptDefense Status=With Defense, Backbone=GPT-4o2026.06 | 87.5 | 67.5 | 5 | |
| Tool FilterDefense Status=No Defense, Backbone=GPT-4o2026.06 | 77.5 | 55 | 5 | |
| System PromptDefense Status=No Defense, Backbone=GPT-4o2026.06 | 75 | 50 | 5 | |
| System PromptDefense Status=With Defense, Backbone=GPT-4o2026.06 | 67.5 | 65 | 0 | |
| Tool-GuardDefense Status=No Defense, Backbone=GPT-4o2026.06 | 67.5 | 62.5 | 5 | |
| Repeat User PromptDefense Status=No Defense, Backbone=GPT-4o2026.06 | 65 | 62.5 | 7.5 | |
| Tool-GuardDefense Status=With Defense, Backbone=GPT-4o2026.06 | 65 | 60 | 2.5 | |
| DriftDefense Status=With Defense, Backbone=GPT-4o2026.06 | 62.5 | 47.5 | 5 | |
| ProagentDefense Status=No Defense, Backbone=GPT-4o2026.06 | 62.5 | 52.5 | 5 | |
| DriftDefense Status=No Defense, Backbone=GPT-4o2026.06 | 60 | 60 | 2.5 | |
| Tool FilterDefense Status=With Defense, Backbone=GPT-4o2026.06 | 57.5 | 45 | 0 | |
| ProagentDefense Status=With Defense, Backbone=GPT-4o2026.06 | 32.5 | 15 | 0 |