Defense and governance
AI guardrails
Guardrails are the runtime checks that constrain what an AI agent can say or do, blocking or reshaping unsafe inputs, outputs, and actions.
Definition
What are AI guardrails?
Guardrails are the enforcement checks that wrap a model or agent at runtime. They span input filtering (catching injection and jailbreak attempts), output filtering (blocking leaked secrets, personal data, or unsafe content), and action controls (constraining which tools an agent may call and with what arguments). Unlike safety training, guardrails are explicit rules you configure and can audit.
For agents, the most consequential guardrails are on actions, because that is where harm actually happens. A guardrail that scores every tool call and blocks a destructive or exfiltrating one is what contains a compromised agent. Effective guardrails are specific, testable, and logged, so you can prove which rule fired and why.
FAQ
Common questions.
Are guardrails the same as model safety training?
No. Safety training shapes the model's tendencies and is probabilistic. Guardrails are explicit, configurable rules enforced at runtime around the model, so they can hard-block a specific unsafe action and be audited afterward.
Related terms
Keep reading.
Govern the actions, not just the vocabulary
AxioRank scores every tool call your agents make for leaked secrets, PII, destructive operations, and prompt injection, checks it against your policy, and proves it in a tamper-evident audit log. Start free, no card.