Guardrails
Filters, monitors, and rule-based constraints layered around a language model to shape acceptable inputs, acceptable outputs, and permitted actions.
Guardrails are the non-model components of an AI system that enforce behavioral limits. The model itself does most of the work through training; guardrails are the belt-and-suspenders layer that catches known failure modes at the edges.
Common categories: input guardrails (reject prompts matching known-bad patterns before they hit the model), output guardrails (filter or transform the model's response before it reaches the user), action guardrails (require confirmation before high-stakes tool calls), and monitoring guardrails (log and alert on suspicious patterns for later review).
Guardrails are usually implemented as separate services: a classifier that scores incoming messages for policy risk, a content filter that scans outputs, a permission system that gates tool use. Some are provided by the model host (Anthropic offers safety filters); others are built by the application.
The tradeoffs are real. Aggressive guardrails frustrate legitimate use (a medical professional getting refused on drug interaction questions). Loose guardrails increase policy risk. Every product picks a point on that curve and defends it.