Atlas

Benchmarks

← All benchmarks

CAISI Cyber Safeguards - No Jailbreak

Safety · 2026-07-17

The no-jailbreak condition of CAISI Cyber Safeguards reports the percentage of responses graded fully compliant with a malicious cyber request. It measures baseline safeguard coverage without adversarial prompting.

Top models (lower is better)

ModelScore
GPT-5.50.0
GLM-5.22.0
Kimi K2.66.0
Opus 4.87.0
DeepSeek-V4-Pro10.0
Loading Atlas data…