# Matimo Guardrails

Guardrails check the content of what an agent receives and says. Policies decide
which actions an agent may take. Guardrails decide whether a response should be
delivered at all. They are separate on purpose.

## What it does

- **Always on.** Every agent has a Default Guardrail on every plan. It blocks
  exposed secrets and high-confidence prompt-injection attempts without any
  setup.
- **Content categories.** Hate, insults, sexual content, violence, misconduct,
  and bias or disparate treatment, each with its own strength for input and
  output.
- **Self-harm handled with care.** A self-harm flag does not produce a flat
  block. It replaces the reply with a message pointing to supportive resources.
- **A fixed safety floor.** A child-safety check is always on and cannot be
  switched off.
- **Sensitive data.** Personal data and secrets can be masked or blocked. You
  can add your own keywords, phrases and patterns.
- **Hold for review.** A flagged response can pause until a person decides.
- **Beyond chat.** The same screening applies to the inputs and outputs of tool
  calls.
- **Measured.** A built-in test panel reports precision and recall per category,
  and an activity log shows what was flagged.
- **Your judge model.** You choose which model judges the content categories.

## What it does not do

- It does not control which tools an agent may call. That is the job of
  Governance policies.
- Custom named guardrails are a plan feature. The Default Guardrail is on every
  plan.

## Who it is for

Anyone putting an agent in front of customers, staff or regulated data.

## Next step

[Contact us](/contact) for a walkthrough of how Guardrails apply to your use
case.