Guardrail types

Reference for the eight guardrails: what each looks for, what you set, and its limits.

Each guardrail has an on and off switch, a choice of what it applies to, a choice of what it does when it fires, and its own settings.

• Applies to: What people send, What the model answers, or Both. Two guardrails have a fixed choice, noted below.
• When it fires: Log only or Block for every guardrail, and Mask as well for the two data protection guardrails.

A guardrail that needs a list checks nothing until you fill it, and cannot be applied while it is on and empty.

Data Protection Filter

Finds personal and sensitive data. Choose the kinds of data to look for, from 31 in six groups:

Group	Kinds
General	Name, phone, email, address, age, username, password, driver ID, license plate, vehicle identification number
Finance	Credit or debit card CVV, expiry and number, PIN, international bank account number, SWIFT code
Technology	IP address, MAC address, URL, cloud access key, cloud secret key
United Kingdom	Unique taxpayer reference number, national insurance number, national health service number
United States	Passport number, social security number, individual tax identification number, bank account number, bank routing number
Canada	Health number, social insurance number

Can Mask.

Custom Data Protection Filter

Finds your own formats, such as account numbers, employee IDs or case references. Add up to 10 patterns, each with:

Field	Limit
Name	Up to 100 characters
Pattern	A regular expression of up to 500 characters, without slashes or flags
What it finds	Up to 1000 characters

Lookahead, lookbehind and backreferences are not supported, and a repeated group that repeats inside itself is refused because it can stall matching. The editor checks the pattern as you type and says what is wrong. Can Mask.

Harmful Content Filter

Finds harmful content in the categories you choose: Hate, Insults, Sexual content, Violence and Misconduct. Sensitivity is Low, Medium or High.

Prompt Injections

Finds attempts to override the AI system’s instructions: jailbreaks, role manipulation, and attempts to extract the system prompt. Always checks what people send. Sensitivity is Low, Medium or High.

Profanity Filter

Finds profanity using a standard word list. No settings beyond the common ones.

Relevance Check

Finds answers that are off topic or not supported by what the model was given. Always checks what the model answers. Relevance needed is from 0 to 0.99: the higher it is, the more relevant an answer must be.

Restricted Topics Filter

Finds subjects you do not want the AI system to discuss. Add up to 30 topics, each with:

Field	Limit
Topic	Up to 100 characters: letters, numbers, spaces and - _ ! ? .
What counts as this topic	Up to 200 characters

Words & Phrases Filter

Finds exact words and phrases, such as competitor names or banned terms. Enter one on each line, up to 10,000, each up to 100 characters.

Controls Altrum enforces

Some risks apply to every AI system and are answered by controls Altrum enforces itself. In assessment results these controls carry an Always applied tag. They are not one of the eight guardrails and have no settings.