85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Glossary

What are AI guardrails?

AI guardrails are checks and limits placed around an AI model or agent that screen its inputs, outputs and actions, and block, correct or escalate anything that breaks a defined policy.

A model will sometimes answer off-topic, leak personal data, follow an injected instruction or take an action it should not. Guardrails are the checks that catch those cases: classifiers on the input, validators on the output, and rules on which tools can run without approval.

Guardrails reduce risk; they do not remove it. Lab guidance in 2026 pairs them with hard limits on what an agent can reach at all.

Types of guardrail

  • Relevance classifiers

    Flag requests that fall outside what the system is for.

  • Safety and moderation checks

    Screen inputs and outputs for harmful content and jailbreak attempts.

  • PII filters

    Detect and redact personal data before it reaches the model or leaves the system.

  • Rules-based checks

    Blocklists, regex, length limits and format checks that run in code.

  • Output validation

    Confirm that an answer matches the schema, cites a source or stays within policy before it is shown.

  • Tool risk ratings

    Rate each tool low, medium or high by reversibility and impact, and require approval for the risky ones.

Guardrails and containment

OpenAI's practical guide to building agents treats guardrails as layered defence, run alongside the agent and able to halt it through tripwires. Anthropic's 2026 work adds that probabilistic checks miss things, so the damage an agent can do should also be capped by deterministic boundaries: sandboxes with file system and network isolation, and credentials kept out of the agent's reach.

Anthropic publishes the limits of its own classifier-based approach: Claude Code's auto mode has a 0.4% false positive rate and a 17% miss rate on overeager actions, and Anthropic says it does not replace careful human review. Google DeepMind's CaMeL design separates a planning model from a quarantined model that reads untrusted data, so injected text cannot change control flow.

How Motiontic's engineers apply it

Motiontic's engineers start from the actions an agent can take, rate each one by impact, and set the boundary before writing any classifier. Guardrails then cover what the boundary cannot, such as personal data in text or an off-policy reply. Each guardrail has its own eval set, so its false positive and miss rates are known.

For regulated work, the guardrails, approvals and logs are designed with the compliance team, as described on the compliance agents page.

Guardrails compared with containment

Guardrails compared with containment
GuardrailsContainment
How it worksChecks content and actions as they happenLimits what the agent can reach at all
ExamplesClassifiers, PII filters, output validatorsSandboxes, network allowlists, scoped credentials
NatureProbabilistic: can miss or over-blockDeterministic: holds regardless of model behaviour
CoversContent quality and policyBlast radius of any failure or attack

Common questions

What are examples of AI guardrails?

Relevance and safety classifiers on inputs, PII redaction, moderation of outputs, schema and policy validation before a reply is shown, and approval requirements on high-risk tools such as refunds or deletions.

Do guardrails stop prompt injection?

They reduce it but do not stop it. The labs say prompt injection is reduced, not solved, which is why guardrails are paired with sandboxes, limited permissions and designs that cap what a successful injection can change.

What is the difference between guardrails and human in the loop?

Guardrails are automatic checks that run on every input, output or action. Human in the loop routes selected steps to a person for approval. Most production agents use both, with guardrails deciding what needs a person.

Are AI guardrails enough for compliance?

Not on their own. Compliance also needs audit logs, access controls, documented approvals and evidence that the controls work, which is why guardrails are tested with evals and recorded with every decision.

Running agents that have to hold up?

Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.