GLOSSARY
Guardrails
The safety layer around an AI system: filters and rules that catch toxic output, blocked topics, prompt attacks and unsafe actions before they reach users.
A model alone is unpredictable by design; guardrails are the product decisions wrapped around it. They include input filters (is this request allowed?), output classifiers (is this response toxic or leaking secrets?), topic boundaries, rate limits, and action checks — an agent may propose “delete the database”, the guardrail layer asks for human confirmation.
Good guardrails are invisible when working and annoying when miscalibrated — every viral screenshot of an AI refusing something harmless is a calibration failure, and every jailbreak is a gap. Vendors expose them as moderation APIs (OpenAI, Cohere) so teams building on LLMs don't have to write their own.