Evaluation profile

WildClawBench Safety & Alignment (OpenClaw harness)

1sub-evals

0.955%total index weight

3components

Within-component eval weight: Human rights 2.24% · Truthfulness 0.458% · Responsible agency 3.67%.

Model score (higher is better)Predicted score

About this eval

Safe execution of long-horizon agent workflows under adversarial and high-risk conditions.

Included in the behavior ranking.

Sub-evals

Measure	Component	Direction	Total index weight	Within-component weight
safety_alignment_score_pctwildclawbench-safety/wildclawbench-safety.csv:safety_alignment_score_pctMeasures whether an OpenClaw agent completes realistic workflows while preserving files, credentials, authority boundaries, injection resistance, and truthful behavior.	human_rights_systemic_harm:0.200\|truthfulness_honesty:0.100\|responsible_agency_control:0.700wildclawbench-safety-alignment	Higher is better	0.955%	Human rights 2.24% · Truthfulness 0.458% · Responsible agency 3.67%

safety_alignment_score_pct

Measures whether an OpenClaw agent completes realistic workflows while preserving files, credentials, authority boundaries, injection resistance, and truthful behavior.

Rank	Model	Value	Provenance
1	claude-opus-4.6	60.04	official
2	kimi-k3	57	official
3	claude-opus-4.8	52	official
4	glm-5	47.36	official
5	glm-5.2	47	official
6	gemini-3.1-pro-preview	45.77	official
7	muse-spark-1.1	42	official
7	qwen3.8-max	42	official
9	qwen3.5-397b-a17b	40.53	official
10	step-3.5-flash	40.41	official
11	grok-4.5	39	official
11	hy3	39	official
13	gpt-5.4	38.49	official
14	gpt-5.6-sol	38	official
15	mimo-v2-pro	37.48	official
16	minimax-m2.7	36.86	official
17	claude-fable-5	36	official
18	kimi-k2.7-code	35	official
19	grok-4.20	33.55	official
19	mimo-v2-flash	33.55	official
21	glm-5-turbo	32.19	official
22	deepseek-v3.2	30.46	official
23	minimax-m2.5	29.8	official
24	kimi-k2.5	26.05	official