Stockertim Dictionary · Artificial intelligence
Guardrail
Also called guardrails, AI guardrails
In short
A guardrail is a rule or check that limits what an AI system may say or do, such as blocking a topic, capping a payment or requiring approval before an action.
- Intermediate
- Concept
- Checked
In detail
Guardrails are the limits placed around an AI system so that its mistakes, and attempts to misuse it, stay within acceptable bounds. Some act on content: filters that block certain topics, remove personal data from outputs or check that a reply cites its source. Others act on actions: spending caps, allow-lists of permitted recipients, approval steps, sandboxes for testing and a stop switch that halts the system.
The most dependable guardrails sit outside the model, in ordinary software the model cannot talk its way past. An instruction in the system prompt saying never pay more than a set amount is a guideline; a check in the payment function that refuses larger amounts is a guardrail. Layering both is common, but the outside check is the one that holds.
Guardrails are designed from the risks of the task. The question is what the worst plausible mistake looks like and which limit would stop or contain it. They are tested like any other feature, including with deliberate attempts to get round them, and reviewed when the system gains new tools or data.
Worked example
Harbour Foods gives a purchasing agent a set of guardrails.
- Orders only from suppliers on the approved list.
- No single order above a fixed limit without a manager's approval.
- No more than ten orders in one day; the agent stops and alerts a person if it reaches the cap.
- A stop switch lets the purchasing lead pause the agent at once.
Even if the agent misreads a request or is misled by an email, the damage is capped and quickly visible.
In practice
Teams list guardrails for every AI feature from its risk assessment, enforce them in code wherever possible, test them with attempts to bypass them and log every time one is triggered.
Watch out
- Instructions in a prompt are guidance; enforceable guardrails sit in the software around the model.
- Too many content filters cause over-refusal, which pushes users towards unofficial tools.
Learn it properly
Official sources
- Guidelines for secure AI system development · National Cyber Security Centre
- Top 10 for large language model applications · OWASP Foundation
- AI Risk Management Framework · National Institute of Standards and Technology
Facts last checked on 3 November 2026.
Education, not advice: the Stockertim Dictionary explains what terms mean. It never recommends buying or selling anything.