Which check should I start with?

Whichever matches a risk you can already name. Most teams start with PII detection or Credentials & Secrets on User input, because both are unambiguous and both produce incidents that are easy to judge.

Can one control use two checks?

No. One check per control. Two detections means two controls.

Why does my control default to Custom (LLM check)?

New controls start there. Change it before configuring anything else — the fields below depend on the check.

Introduction

The check is the first of a control's three axes — what it detects. Everything else about a control follows from it, including which actions you are allowed to choose.

You pick the check from one searchable list. Most entries are shipped and configured by UnifyApps; Custom (LLM check) is the entry where you describe the check in plain language.

Where to find it

AI Governance > Controls > New Control > Check type — "the risk area this control governs".

Access to AI Governance is granted through platform roles. Ask your workspace administrator, or see Settings > Governance > Roles for how permissions are assigned.

The checks

Check	Detects	How
`PII detection`	Personal data — emails, phone numbers, US SSNs, card and bank numbers, and more (not names)	Classifier
`Credentials & Secrets`	Secrets — API tokens, passwords, private keys	Classifier
`Prompt Injection`	Instructions hidden in content that try to hijack the agent	Classifier model you choose
`Content Filter`	Common safety categories — hate, violence, sexual, self-harm, and more	Classifier model you choose, with per-category thresholds
`Denied Topics`	Named topics the agent must not discuss	LLM classifier
`Custom terms`	A primary term, with optional related terms of which any one must appear	Term matching with context
`Regex pattern`	A raw pattern matched against the input	RE2 regular expression
`Blocked Tools`	Tools the agent may never invoke, regardless of context	Tool-name rules
`Custom (LLM check)`	Whatever you describe in plain language	A model you choose, with a prompt you write

Choosing between them

Three questions settle it in almost every case.

Is the thing you are detecting a known category? — Personal data, secrets, injection attempts and safety categories all have purpose-built checks. Use them — they are better tested than a pattern you write. PII detection and Credentials & Secrets add no model cost; Prompt Injection and Content Filter call a model you choose.
Can you express it exactly? — A pattern with a stable shape — an internal reference format, an account number — is a Regex pattern. Deterministic, free to evaluate, and no false positives beyond what your pattern actually says.
Does it need judgement? — Then Denied Topics if it is subject-shaped, or Custom (LLM check) if it is not. Both call a model on every evaluation.

"Reach for the cheapest check that can actually answer the question. An LLM check is the fallback, not the default — even though it is what a new control starts on."

Two consequences of the choice

Cost — A pattern or term check costs nothing per input. Four checks call a model and add governance spend: Custom (LLM check), Denied Topics, Prompt Injection and Content Filter — each has a model setting, and makes a model call every time it runs, on every input it reads. See Where Spend Comes From.
Available actions — Mask acts on a matched span, so only checks that report where the match occurred can offer it. A check answering yes-or-no can flag or block only. Custom (LLM check) is in that second group.

The list can change

Prebuilt checks are resolved when the editor opens rather than being fixed in a release, so your tenant's list is the authority. If a check appears here that you cannot see, or you can see one this page does not list, trust the picker.

Check Types