LLM Guardrails.

Mechanisms that constrain what an LLM-based application can input or output, enforcing safety, security, and business rules around the underlying model. It belongs to the AI & ML Security category of cybersecurity.

Mechanisms that constrain what an LLM-based application can input or output, enforcing safety, security, and business rules around the underlying model.

Guardrails are the policy layer of LLM applications. They include classifiers and rule-based filters for prompt-injection or jailbreak attempts, topic, persona, and tone controls, output schema validation, PII or secret scrubbing, refusal handling, citation requirements, and tool-call constraints. Implementations range from open-source frameworks such as NVIDIA NeMo Guardrails, Guardrails AI, and Microsoft's Presidio, to vendor APIs like OpenAI Moderation or Anthropic's safety endpoints, to bespoke logic inside agent frameworks. Guardrails complement model-internal alignment, LLM firewalls, and MLSecOps practices. They should be testable, versioned, and continuously validated by red teaming, since attackers focus on finding the gap between guardrails and model behaviour.

Defences for LLM Guardrails typically combine technical controls and operational practices, as detailed in the full definition above.

Common alternative names include: AI guardrails, Generative AI guardrails.