Evaluation profile

ToolPrivacyBench

2sub-evals

1.25%total index weight

2components

Within-component eval weight: Human rights 6.86% · Responsible agency 1.48%.

Model score (lower is better)Predicted score

About this eval

ToolPrivacyBench

Included in the behavior ranking.

Measure	Component	Direction	Total index weight	Within-component weight
private_mt_poitoolprivacybench/toolprivacybench.csv:private_mt_poiMeasures private mt poi in the cited first-party or paper evaluation.	human_rights_systemic_harm:0.700\|responsible_agency_control:0.300toolprivacybench	Lower is better	0.626%	Human rights 3.43% · Responsible agency 0.738%
public_mt_poitoolprivacybench/toolprivacybench.csv:public_mt_poiMeasures public mt poi in the cited first-party or paper evaluation.	human_rights_systemic_harm:0.700\|responsible_agency_control:0.300toolprivacybench	Lower is better	0.626%	Human rights 3.43% · Responsible agency 0.738%

Measures private mt poi in the cited first-party or paper evaluation.

Rank	Model	Value	Provenance
1	gemini-3.5-flash	19.19	official
2	claude-opus-4.7	20.31	official
3	gpt-5.5	20.39	official
4	doubao-seed-2.0-lite	22.99	official
5	minimax-m2.7	24.47	official
6	deepseek-v4-flash	27.33	official
7	qwen3.6-plus	27.46	official
8	kimi-k2.5	27.74	official
9	glm-5.1	28.04	official

Measures public mt poi in the cited first-party or paper evaluation.

Rank	Model	Value	Provenance
1	kimi-k2.5	15.81	official
2	gpt-5.5	16.75	official
2	minimax-m2.7	16.75	official
4	claude-opus-4.7	17.83	official
5	doubao-seed-2.0-lite	18.78	official
6	deepseek-v4-flash	18.99	official
7	qwen3.6-plus	19.25	official
8	gemini-3.5-flash	19.86	official
9	glm-5.1	22.56	official