Atlas

Benchmarks

← All benchmarks

Aikido Known-CVE Detection (Pass@3 Union)

Code · 2026-07-16

Aikido Known-CVE Detection evaluates whether tool-equipped models can rediscover 26 known vulnerabilities from the GitHub Advisory Database across real repositories. Investigator agents are pointed to each vulnerable code snippet, and the primary score is the share of CVEs surfaced in at least one of three independent runs (a pass@3 union).

Top models (higher is better)

ModelScore
GPT-5.6 Terra88.5
GPT-5.6 Sol88.5
GPT-5.6 Luna84.6
GPT-5.584.6
GPT-5.4 Mini80.8
Grok 4.576.9
GPT-5.4 Nano69.2
Opus 4.869.2
Gemini 3.1 Pro Preview65.4
GLM-5.261.5
Opus 4.757.7
Haiku 4.553.8
Gemini 3.5 Flash53.8
Loading Atlas data…