Definition
AI guardrails is AI guardrails are the rules, filters and checks placed around an AI model or agent to keep its inputs, outputs and actions within safe and intended limits.
AI guardrails are the controls that keep an AI system inside acceptable bounds. They are not one feature. They are a layered set of protections applied before, during and after a model does its work, from blocking harmful requests to limiting what an agent is allowed to touch.
Guardrails can sit in several places:
Teams build these layers because no single one is reliable alone. A model can be coaxed into ignoring its training, a filter can miss a new phrasing, and an agent can misread an instruction. Stacking checks means one failure does not become an incident.
If you build with AI, guardrails are what make a demo safe to put in front of customers. Good ones are specific to your use. A medical assistant needs rules about advice. A support bot needs limits on refunds. A coding agent needs limits on which folders it can change.
If you only use AI, guardrails explain why a model sometimes refuses a harmless request. Rules tuned to prevent harm will occasionally block something legitimate, and teams tune them with that tradeoff in mind.
Guardrails reduce risk and do not remove it. Determined users probe them, and researchers do this on purpose through red teaming to find gaps before attackers do. Rules also need upkeep as models, tools and misuse change.
Imagine an assistant that can read your email and draft replies. Sensible guardrails would let it read and draft, block it from sending without your approval, hide sensitive attachments from the model, and log everything it did. If a malicious email tries to instruct the assistant to forward your files, the action limit stops it even if the model is fooled.
Hiding a signal in generated content so it can be detected later, as in AI watermarking, is another safeguard, though it works after content exists rather than limiting what the model does.
Layered rules, filters and permission limits that keep an AI model's inputs, outputs and actions within safe and intended bounds.
Guardrails tuned to block harm can occasionally catch legitimate requests, so teams balance caution against usefulness.
CrowdStrike says a likely Chinese-speaking, financially motivated actor ran ARTEX, an open-source agentic pentest tool, against South Korean financial firms from late September, and cites industry reports of breaches at two banks. The model under it was swappable, and that is the warning.