AI Guardrails
The safety layer around an AI system — filters, policies, validation, and monitoring — that constrains what a model can input, output, or do before actions reach users or downstream systems.
AI guardrails are the practical controls that keep a model’s behaviour within acceptable bounds. They sit outside the model itself and apply at defined checkpoints: on the way in, on the way out, and around any actions the system takes.
Typical guardrail components:
- Input filters: block or sanitise disallowed requests and strip known injection patterns.
- Output classifiers: detect harmful, confidential, or off-policy content before it is shown.
- Schema and policy validation: enforce that structured outputs and tool calls are well-formed and permitted.
- Action gates: require human approval for irreversible operations such as payments, deletions, or external messages.
- Monitoring: log decisions and flag anomalies for review.
Guardrails are distinct from alignment, which concerns a model’s learned values and tendencies. A well-aligned model can still be manipulated or make mistakes, so production systems rely on both. As agents gain autonomy, guardrails increasingly focus on limiting the blast radius of any single failure rather than trying to prevent every possible one.