What are guardrails, and what types are there?

30-second answerSay your answer out loud first, then reveal.

Guardrails are controls that constrain an AI system's inputs, outputs and actions so it behaves safely and within policy, regardless of what the model would do on its own. Input guardrails filter or transform what reaches the model (PII redaction, injection detection, topic restriction). Output guardrails check what the model produces (toxicity, PII leakage, groundedness, format, policy). Action guardrails control what an agent can do (permissions, approvals, rate limits). They can be rules, classifiers, LLM-based checks, or architectural constraints.

Type	Examples	Implementation
Input	PII redaction, jailbreak/injection detection, off-topic filter, length limits, language detection	Regex, NER, classifiers, small LLM
Output	Toxicity, PII leak, hallucination/groundedness, schema validation, competitor mentions, brand tone	Classifiers, validators, LLM judges
Action / tool	Allow-listed tools, parameter validation, spending limits, approvals, sandboxing	Code-level policy enforcement
Dialogue / flow	Scripted flows for sensitive topics, escalation to humans	Rails frameworks, state machines
System-level	Rate limits, abuse detection, monitoring, kill switches	Infrastructure

Key principle: guardrails in code and architecture are reliable; guardrails in prompts are guidance the model may not follow. Use both, but enforce critical policies in code.

Trade-offs: each guardrail adds latency and cost and can cause false positives (over-blocking). Tune thresholds with evals (Q28).

PreviousWhat is error analysis, and why is it called the most important eval activity?

You understood something today that you didn't yesterday.

What are guardrails, and what types are there?

Related