Agent Guardrails

Overview

A single LLM call cannot be trusted to police itself. Production agents need layered defenses: input filters, output filters, tool allow-lists, and behavioral policies. This skill picks the libraries and wires them in at the right points.

When to use

• Agent reads untrusted input (web pages, emails, user uploads)
• Agent has tools that touch real systems (DB, API, money, send messages)
• Compliance requires PII redaction or content policy enforcement
• You've seen prompt injection or data exfiltration attempts in logs
• LLM-as-judge alone has been bypassed in red-team tests

Threat model (2026)