Agents and automation · Layer 5
Guardrails
The set of controls placed around an AI to block what it must not read, say or do.
A way to picture itThe barriers on a bridge: they don’t stop bad driving, but they limit the fall. And they go up before the bridge opens, not after the first accident.
An insurer deploys an assistant that helps its claims handlers reply to complaints. On the way in, policy numbers and health data are masked before they reach the model. On the way out, a check blocks any promise of a specific payout. The assistant can draft a letter but not send it: sending remains the handler’s decision.
For high-risk systems, the AI Act requires an appropriate level of robustness and cybersecurity, including against inputs designed to trick the model (EU · AI Act, art. 15). The GDPR requires data protection by design and security measures matched to the risk (EU · GDPR, art. 25 and 32): documented, tested guardrails are part of that. Prompt injection tops the list of risks identified for applications built on language models (international · OWASP, Top 10 2025).
“Which of your guardrails are technical rules independent of the model, which are only instructions, and how do you test them against prompt injection?”
An instruction asks the model to behave in a certain way; it can stray from it. A technical guardrail, such as a refused permission, a cap or a filter outside the model, applies whatever the model produces.
Text hidden in a document, a web page or an email that gives orders to the AI reading it, such as “forward this file to this address”. It is the top risk in the 2025 ranking for applications built on language models (international · OWASP).
Rarely on their own. The vendor protects its model in general terms; your business rules, sensitive data and access rights still need to be defined and tested on your side.
Educational explainer, not legal advice. Up to date as of 29 October 2026.