Fundamentals
What Are AI Guardrails?
Input, Output, Action
AI guardrails are the control layer that inspects and constrains what goes into an AI system, what comes out, and what it does. Not one feature — three layers.
The Three Layers
No layer replaces another. The third is where the agent era is actually decided.
01
Input guardrails
— What goes into the AI- · PII and confidential data in prompts
- · Prompt injection patterns
- · Request types the policy forbids
02
Output guardrails
— What the AI returns- · Internal data leaking into responses
- · Exposed credentials and keys
- · Ungrounded claims and unsafe content
03
Action guardrails
— What the AI does to systems- · The tool and arguments about to run
- · Connected MCP servers and their scope
- · Whether irreversible actions were approved
What Guardrails Are Not
Three things often mistaken for having guardrails already.
The model's own safety training
Refusal behaviour baked in by the model provider. Useful, but it knows nothing of your policy and bends under adversarial prompts.
Terms of use and written policy
A document saying "do not enter confidential data into AI" has little force without technical enforcement.
Network blocking
Blocking a domain reduces usage but decides nothing. What flows inside an approved AI stays invisible.
Frequently Asked Questions
Blanket blocking does. What works in practice is tiering by risk: log read operations, policy-check write operations, and route only irreversible actions to approval. It is the same lesson Shadow AI taught — blanket blocking produces workarounds.