Atlas

Benchmarks

← All benchmarks

CAISI Harmful Biology and Violent Activities - Best Public Jailbreak High Detail (2025)

Safety · 2025-09-30

This diagnostic reports the share of responses graded as highly detailed on CAISI's 65 harmful-biology and violent-activities requests under the most effective of 17 public jailbreaks selected for each model on held-out data. Detail does not measure factual accuracy or utility.

Top models (lower is better)

ModelScore
DeepSeek-V3.170.0
Loading Atlas data…