Atlas

Benchmarks

← All benchmarks

Aikido Known-CVE Detection (32 CVEs)

Code · 2026-10-22

Aikido's October 2026 run of its known-CVE rediscovery benchmark on an expanded set of 32 recently disclosed vulnerabilities, using the AI Code Analysis harness with investigator agents pointed to the vulnerable code. The headline is pass@3 union recall: the share of the 32 CVEs found in the union of three runs. Not comparable with the 26-CVE July set.

Top models (higher is better)

ModelScore
GPT-6 Astra90.6
DeepSeek V4.1 Flash 101787.5
DeepSeek V4 Pro 081387.5
Loading Atlas data…