Guardrails
AI agent guardrails are the checks that stop an agent from taking an action, or passing on content, that it should not. They sit outside the model, so they apply whatever the model decides.
In a run
A supplier attachment carries a hidden instruction to send a drawing to a new address. The instruction is stripped before the agent reads it, the send is refused as outside permission, and both decisions go on the record.
Our word for it
Read at the door; checked before it leaves.
Two kinds of check
Some guardrails act on content. Documents are read at the door before an agent sees them. Answers are checked, and redacted where needed, before they leave.
Others act on actions. A call the access policy does not cover is refused. A prompt can talk a model into trying something. It cannot widen the credential.
Steps that wait for a person
Some actions are allowed, but not alone. A delivery date change waits for the buyer. Some kinds of step wait for a person by default, such as widening a permission or keeping data beyond the job. You add the steps that matter in your business.
Guardrails that leave a trace
A refusal that nobody can see later does not help an auditor. On Ares, the filters that fired and every refusal, with its reason, are on the record. You can see which filter fired on which document.
Where it stops
Filters cover stated patterns and languages. The permission check does not depend on them: a call outside the credential is refused either way.
See a hidden instruction caught in the underwriting scenario.
Related terms
- Content inspectionContent inspection checks what flows into and out of an AI agent against stated rules.
- Refusal recordAn AI agent refusal record is the entry written when an AI agent asks to do something and is not allowed to.
Tell us about one workflow.
We will say whether an agent should do it, and what the first step costs.
A senior member of the team replies within two business days.