Aikido Known-CVE Detection (Single-Run Average)
Code · 2026-07-16
The single-run companion to Aikido Known-CVE Detection reports mean recall across three independent audits of the same 26 known vulnerabilities. It measures average one-run consistency rather than the pooled union used by the primary benchmark and is retained as a diagnostic metric.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.6 Terra | 83.3 |
| GPT-5.6 Sol | 78.2 |
| GPT-5.6 Luna | 74.4 |
| GPT-5.5 | 73.1 |
| Grok 4.5 | 66.7 |
| GPT-5.4 Mini | 64.1 |
| GPT-5.4 Nano | 57.7 |
| Opus 4.8 | 55.1 |
| Gemini 3.1 Pro Preview | 47.4 |
| GLM-5.2 | 46.2 |
| Opus 4.7 | 44.9 |
| Gemini 3.5 Flash | 44.9 |
| Haiku 4.5 | 41.0 |