← Evals

Evaluation profile

MT-JailBench CrescendoX

1sub-evals
0.0213%total index weight
1components

Within-component eval weight: Misuse resistance 0.213%.

Model score (higher is better)Predicted score

About this eval

Multi-turn jailbreak resistance under fixed CrescendoX conditions.

Included in the behavior ranking.

Sub-evals

MeasureComponentDirectionTotal index weightWithin-component weight
safety_scoremt-jailbench/mt-jailbench.csv:safety_scoreMeasures resistance to the fixed CrescendoX multi-turn jailbreak protocol.ordinary_harm_misuse_resistance:1.000harmbenchHigher is better0.0213%Misuse resistance 0.213%

safety_score

Measures resistance to the fixed CrescendoX multi-turn jailbreak protocol.

RankModelValueRelative performanceProvenance
1claude-haiku-4.581.76official
2claude-sonnet-4.575.47official
3gpt-561.29official
4qwen3.5-plus47.8official
5nemotron-3-super-120b-a12b44.03official
6gemini-3-pro-preview36.48official
7qwen3.5-flash34.59official
8llama-3-8b-instruct28.93official
9gemma-4-26b-a4b-it26.58official
10llama-4-maverick22.64official
11gemma-4-e4b21.38official
12grok-4.1-fast18.87official
13deepseek-v3.217.61official
14qwen3-4b-2507-instruct16.35official
15gpt-4o15.72official
16gemini-3-flash-preview13.84official
17llama-3-70b-instruct11.32official
17nemotron-3-nano-30b-a3b11.32official
17qwen3-30b-a3b-instruct11.32official
20llama-4-scout10.06official
21ministral-3-8b0.63official