Research2026-08-21
Security vendor Aikido published results from running ten AI models three times each against 32 recently disclosed software vulnerabilities, asking them to rediscover the flaws from source code. DeepSeek V4 Pro found the most when results from all three runs were combined, beating the publicly available closed models and costing far less. The open models also produced the most false leads, and Aikido notes they still need a carefully built agent setup around them.
What changed
An earlier version of the same test used known CVEs and fewer models, and closed frontier models led on coverage.
What it unlocks
Choosing cheaper models run several times over one expensive pass when hunting for security flaws in source code.
- DeepSeek V4 Pro: 28 of 32 pooled
- 3 Pro runs ≈ $295
- 3 Flash runs ≈ $108, 24 found
- single frontier pass ≈ $450-$590
What you need to act on it
- an agent setup that explores code and filters false leads
- API access to the models tested
- aikido.dev2026-08-21