Glossary · AI & Development
Guardrails
Controls applied around or inside an AI system to constrain its outputs and prevent harmful, incorrect or out of scope responses.
Browse all definitionsIn detail
Guardrails is a broad term for the technical and procedural controls that bound what an AI system can do. Input guardrails screen prompts before they reach the model (blocking injection attempts, off topic queries or personally identifiable information). Output guardrails validate model responses before the system returns or acts on them (checking format, detecting hallucinations, blocking harmful content). Architectural guardrails limit what tools an agent can call and what data it can access. Guardrails can be implemented as classifier models, rule based filters, JSON schema validation, human in the loop approval flows, or rate limits on consequential actions.
Sources & further reading
Check the source, not just the summary
- OWASP Top 10 for LLM Applicationsgenai.owasp.org
Apply the definition
Want to talk through how this applies to your business?
Start with the decision in front of you. We will help map the fit.