Opinion: security teams should add a deterministic second opinion to AI scans

Checkmarx research shows that running the same scan five times on identical code yields matching findings only about a third of the time, highlighting inconsistency in frontier AI scanners. The industry average F1 score sits around 0.20, but Checkmarx claims its approach delivers three to four times the accuracy of a standard scanner. Deterministic, rules‑based analysis provides consistent, auditable results, while frontier models add reach for novel patterns. Together they can cut costs – scanning a one‑million‑line repository costs more than a thousand dollars – and reduce the “token tax” that Uber, which spends $3.4 billion a year on R&D, reportedly exhausted its 2026 AI budget within months of rolling out AI coding tools.