Evaluation profile
BioSecBench-Refusal
1sub-evals
0.685%total index weight
2components
Within-component eval weight: Benign helpfulness 6.1% · Misuse resistance 0.758%.
Model score (higher is better)Predicted score
About this eval
Calibrated refusal across legitimate biological work and concealed biosecurity hazards.
Included in the behavior ranking.
Sub-evals
| Measure | Component | Direction | Total index weight | Within-component weight |
|---|---|---|---|---|
| balanced_refusal_scorebiosecbench-refusal/biosecbench-refusal.csv:balanced_refusal_scoreMeasures whether the model helps with routine biological research while refusing concealed requests that could enable biological harm. | benign_helpfulness_non_overrefusal:0.500|ordinary_harm_misuse_resistance:0.500biosecbench-refusal | Higher is better | 0.685% | Benign helpfulness 6.1% · Misuse resistance 0.758% |
balanced_refusal_score
Measures whether the model helps with routine biological research while refusing concealed requests that could enable biological harm.
| Rank | Model | Value | Relative performance | Provenance |
|---|---|---|---|---|
| 1 | gemini-3.5-flash | 0.5041 | official | |
| 2 | claude-sonnet-4.6 | 0.4304 | official | |
| 3 | claude-opus-4.7 | 0.4259 | official | |
| 4 | gpt-5.4 | 0.4254 | official | |
| 5 | claude-opus-4.6 | 0.3987 | official | |
| 6 | gemini-3.1-pro-preview | 0.3931 | official | |
| 7 | claude-opus-4.8 | 0.3633 | official | |
| 8 | gpt-5.5 | 0.2869 | official | |
| 9 | grok-4.3 | 0.1092 | official | |
| 10 | grok-4.20 | 0.02854 | official |