WHITEBOXX / GLOSSARY

AI guardrails: the limits
on what an agent can do.

AI guardrails are controls that limit what an AI system can receive, produce and do. They include filters on inputs and outputs, rules about which tools and data an agent may use, and limits on spending. For agents, the limits on actions matter most.

By Gautam Mehra · Published · Updated

What are AI guardrails?

For a chat assistant, guardrails mostly filter text: blocking harmful requests or removing personal data from answers. For an agent that can query systems and take actions, guardrails also decide what the agent is allowed to touch.

Guardrails work best in layers. No single filter catches everything, so a system combines checks before, during and after each step.

TYPES

Four types of guardrail.

TYPES
TypeWhat it controlsExample
InputWhat reaches the modelDetecting instructions hidden in a document the agent reads
OutputWhat the system returnsChecking an answer against its sources before it is shown
ActionWhat the agent may doRead access to the CRM, but no permission to send emails
BudgetHow much it may useA limit on model calls, time or cost for each task

Which risks do guardrails address?

The 2025 edition of the OWASP Top 10 for LLM Applications lists the most common security risks. Two matter most for agents. Prompt injection is when text the agent reads, such as a web page or an email, contains instructions that change its behaviour. Excessive agency is when an agent has more tools, permissions or autonomy than its task needs.

The simplest defence against both is to give each agent only the tools and data its task requires.

TESTING

How to test your guardrails.

  1. 01

    List what each agent can reach

    Write down every tool, system and dataset, and the permission it has.

  2. 02

    Remove what the task does not need

    Least privilege applies to agents as it does to people.

  3. 03

    Try to break it

    Give it documents with hidden instructions and requests outside its task.

  4. 04

    Record refusals as well as actions

    Keep a record of what the agent tried to do, not only what it did.

How Evidence OS limits agents.

Each execution-agent assignment in Evidence OS runs with a defined set of tools and a budget, and the run record shows those limits alongside what the agent did.

See execution agents