Small open-weight models for the trust layer of enterprise AI agents: guardrail enforcement, scam and fraud detection, and PII identification across C