Guardrails

Rules enforced around the model — checking inputs and outputs, limiting tools — instead of hoping it behaves.

Security path · 3 of 5

What it is

Guardrails are controls outside the model: filters on inputs and outputs, allow-lists of tools, limits on data access, redaction of personal data, rate and cost limits. They work at request time and don't depend on the model following instructions.

Example

Before an answer is shown, a filter removes payment card numbers that slipped into the text.

What can go wrong

Guardrails written only as instructions in the prompt ("never reveal…") can be talked around.

How to secure it

Enforce guardrails in code and infrastructure, test them with red teaming, and log what they block.

Related service: AI Security