Review Guardrails

Guardrails are technical, operational, or policy controls intended to keep an AI system's inputs, outputs, actions, or uses within defined boundaries.

Guardrails may filter content, restrict tools, limit permissions, require approval, monitor behavior, or stop work when conditions fall outside an accepted range.

A model's behavior alone is not a complete safety system. Effective controls need clear objectives, testing, authorized alternatives, monitoring, and ownership.

A procurement assistant may draft a purchase request but cannot submit it until an authorized employee reviews the vendor, amount, and account code.

Guardrails are not guarantees and should not be treated as a substitute for sound architecture, limited access, human ownership, or incident response.

Guardrails define and enforce the operating lane.

* [[guardrail-displacement|Guardrail Displacement]] * [[guardrail-inversion|Guardrail Inversion]] * [[human-ownership|Human Ownership]] * [[red-teaming|Red Teaming]] * [[objective-escape|Objective Escape]]

* [NIST, *Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile* (NIST AI 600-1)](https://doi.org/10.6028/NIST.AI.600-1) — risk-management actions and controls for generative-AI systems across the lifecycle. * [UK AI Security Institute, *Principles for Safeguard Evaluation*](https://www.aisi.gov.uk/blog/principles-for-safeguard-evaluation) — official framework for evaluating safeguards as system controls rather than guarantees.

* Test controls through [[red-teaming|Red Teaming]] and watch for [[guardrail-displacement|Guardrail Displacement]].