Research

Reaching Frontier Parity in Security Code Remediation with an On-Premise 22B Model

We fine-tune Codestral-22B on 4,811 validated vulnerability fix triples using SFT + DPO in 4-bit QLoRA. Results show parity with Claude Opus 4.8 on CyberSecEval remediation and a non-significant gap on SecurityEval.

We Scanned 1,072 Vibe Coded Apps, 98% Had Security Flaws!

We crawled 65,643 URLs and scanned 1,072 Supabase backed vibe coded apps. 98 percent had security issues and 16 percent had critical flaws. Learn what is failing and what to fix first with RLS.

See all articles

Book a demo

See how our solution empowers teams to grow their security maturity and to code securely & efficiently.

Demo illustration