Blog
- 5 min readResearch
Open-weight vs frontier models: the economics of an AI security harness
Model choice is only part of the cost equation. An efficient security harness uses lower-cost models for broad discovery, then reserves expensive models for verification and the hardest decisions. Differential AI SAST lowers costs further by scanning new code instead of the whole repository. In our AI-assisted development scenario, it reduces the workload from 4.86x to 0.06x, or over 98%.
- 13 min readResearch
What AI finds in code that pattern matching cannot
Three AI harnesses produced 1,318 findings across 4.8 million lines of production code. A third sat in authorization, IDOR and race classes that conventional SAST has no rule for. Almost none reproduced between runs, and none had been verified.