Illicit Task Completion on AgentHarm English prompts
72.7AgentHarm Score (AHS)STING
Evaluation Results
| Method | Links | |
|---|---|---|
| STINGTarget Agent Model=Qwen3-Next, Smax=102026.02 | 72.7 | |
| STINGTarget Agent Model=Qwen3-Next, Smax=52026.02 | 67.8 | |
| STINGTarget Agent Model=DeepSeek-V3.2, Smax=102026.02 | 61.8 | |
| STINGTarget Agent Model=DeepSeek-V3.2, Smax=52026.02 | 57.2 | |
| STINGTarget Agent Model=Gemini 3 Flash, Smax=102026.02 | 50.9 | |
| STINGTarget Agent Model=Gemini 3 Flash, Smax=52026.02 | 47.6 | |
| Single-turn promptingTarget Agent Model=Gemini 3 Flash2026.02 | 45.9 | |
| Single-turn promptingTarget Agent Model=Qwen3-Next2026.02 | 35.1 | |
| STINGTarget Agent Model=GPT-5.1, Smax=102026.02 | 34.1 | |
| STINGTarget Agent Model=Claude Sonnet 4.5, Smax=102026.02 | 32.3 | |
| Single-turn promptingTarget Agent Model=DeepSeek-V3.22026.02 | 31.2 | |
| STINGTarget Agent Model=GPT-5.1, Smax=52026.02 | 29.7 | |
| STINGTarget Agent Model=Claude Sonnet 4.5, Smax=52026.02 | 28 | |
| X-TeamingTarget Agent Model=Qwen3-Next, Smax=52026.02 | 27 | |
| Single-turn promptingTarget Agent Model=GPT-5.12026.02 | 24.3 | |
| Single-turn promptingTarget Agent Model=Claude Sonnet 4.52026.02 | 16 | |
| X-TeamingTarget Agent Model=DeepSeek-V3.2, Smax=52026.02 | 15.1 | |
| X-TeamingTarget Agent Model=Gemini 3 Flash, Smax=52026.02 | 13.8 | |
| X-TeamingTarget Agent Model=GPT-5.1, Smax=52026.02 | 5 | |
| X-TeamingTarget Agent Model=Claude Sonnet 4.5, Smax=52026.02 | 2.2 |