LLM code security review

Haiku labels three of eight patched twins as reachable vulnerabilities

A Kaggle-challenge benchmark called ART scores models on whether they still flag a function after the fix is applied. On eight synthetic pairs, the difference between price tiers showed up only on the patched half.