Guardrails and Banned URLs
This guide covers how to narrow what is passed to the AI, and where agents can go.
What to decide on which screen
Each thing you want to narrow is decided on its own screen.
| What to narrow | Screen | What you decide |
|---|---|---|
| Content passed to the AI | “AI Agent Restrictions” > “Guardrails Settings” | Prompt attack detection, harmful content detection sensitivity, personal information masking, execution time limit |
| Instructions to the AI | “Prohibition Instructions” in “AI Agents” > “Common AI Agent Instructions” | Text passed to the AI as an instruction, stating security policy and what is forbidden |
| Tool operations (project runs only) | “Project Management” > “Execution rules” | Allowing or forbidding operations. They apply only to project step runs and cycle rounds, not to ordinary conversations (Execution Rules Guide) |
| Where agents go | “AI Agent Restrictions” > “Banned URL” | Patterns for destinations to stop |
| Traffic from widgets in agents' answers | “AI Agent Restrictions” > “Widget Connection Restrictions” | Domains widgets (small interactive screens in answers) must not connect to |
Prohibition Instructions are text passed to the AI as instructions, and are also used to check requests when they are received. Requests judged to violate them are not executed. Because the AI makes that judgment, for anything you need stopped for certain, do not rely on the instructions alone; stop it with Guardrails Settings or banned URLs. In project runs, execution rules can stop it too.
LLM Gateway guardrails are a separate setting (“LLM Gateway Management” > “LLM Gateway Guardrails”).
Guardrails Settings
Under “AI Agent Restrictions” > “Guardrails Settings”, decide these five things. The defaults are:
| Item | Choices | Default |
|---|---|---|
| Prompt Shield Detection | On / Off | On |
| Moderation Detection Sensitivity Level | Disabled, Recommended, Lenient, Most Lenient | Recommended |
| Personal information masking | On / Off | On |
| Execution time limit | 5, 15, 30, 60, 90 or 120 minutes / No limit | No limit |
| DLP Integration | Off, Log only, Stop on violation | Off |
Prompt Shield Detection
Automatically detects and stops input that tries to hijack the AI's instructions.
Moderation Detection Sensitivity Level
Content is rated for severity in four categories (hate, sexual, violence and self-harm), and you choose the severity from which it is stopped. Other categories, such as illegal acts, are not checked by this setting.
| Severity | What it looks like (examples) | Recommended | Lenient | Most Lenient |
|---|---|---|---|---|
| 0 | Harmless | — | — | — |
| 2 | Violent scenes in fiction, sexual terms in medical contexts, people describing their own self-harm and asking for help, strong insults | Stopped | — | — |
| 4 | Insulting or demeaning people or groups, graphic violence or how to commit violence, explicit sexual activity, words that encourage self-harm | Stopped | Stopped | — |
| 6 | Terrorist or extremist propaganda and attack planning, calls for violence against groups, sexual exploitation of children, instructions for suicide | Stopped | Stopped | Stopped |
With “Disabled”, nothing is stopped.
- With Recommended, business documents can also be stopped when they involve stories, medical topics, counselling and the like. If content your work needs is being stopped, review it in the direction of loosening
- The severity of a stopped item can be seen in the details of “Guardrail Alerts”, as the same 0, 2, 4 or 6
Personal information masking
When it is on, personal information found in requests and instructions, and in external content an agent reads, is masked before it is passed to the AI. The AI's answers and deliverables are not masked.
What is masked:
- People: names, ages
- Contact: email addresses, phone numbers, addresses
- Financial: credit card numbers, bank account numbers
- National IDs: national identification numbers of each country (such as My Number, driver's licences and passports)
It also applies to the instructions of Project steps. How instructions written by people (such as an agent's goal) are treated is covered in the Design Guide.
Sensitive terms and exclusions
You can add to, or take away from, what personal information masking hides (the list above) to fit your tenant.
| What you want | Register | Example |
|---|---|---|
| Also mask things not on the list (internal project names, employee numbers, API keys and so on) | Sensitive term | Replace “Project Phoenix” with “[Project name]” before passing it to the AI |
| Stop masking words that are mistaken for personal information | Exclusion | Do not mask “Phoenix”, a product name mistaken for a person's name |
Both are registered in the personal information masking section of “Guardrails Settings”. While personal information masking is off, neither applies.
Sensitive terms
What a sensitive term matches is replaced with “[display name]” before it is passed to the AI. Register a word or a regular expression. A word is a case-insensitive partial match.
| Item | Details |
|---|---|
| Type | Term / Regular expression |
| Value | The word to mask (for example Project Phoenix) or a regular expression (for example EMP-\d{6}). Words up to 100 characters, regular expressions up to 500 |
| Display name (required) | Shown as “[display name]” where text is replaced (for example: Project name). Up to 50 characters. [, ] and control characters cannot be used |
| Enabled | Turned off, the entry stays registered but no longer matches |
- Up to 200 can be registered
- To mask strings with a fixed shape, such as passwords or API keys, register a regular expression
- Common regular expression syntax (
\d,[A-Z],{6},|and so on) works as is. The syntax follows RE2, the same as REGEXMATCH in Google Sheets (RE2 syntax). What cannot be used: lookahead and lookbehind ((?=…)and so on), backreferences (\1and so on),\C(which matches a single byte), and expressions that match the empty string (such asa*on its own) - With “Test regular expression” on screen, run it against sample text (up to 2,000 characters) and see what it matches (up to 50 matches)
Exclusions
A string found as personal information is left unmasked only when it is exactly the same as an exclusion. Case is ignored. A partial match does not count.
| Exclusion | String found | Result |
|---|---|---|
| Phoenix | Phoenix | Not masked |
| Phoenix | phoenix | Not masked |
| Phoenix | Phoenix Tanaka | Masked |
- Exclusions are words only (regular expressions cannot be used). Up to 100 characters, up to 200 entries
- Exclusions do not apply to sensitive terms. A word masked by a sensitive term is masked even if the same word is an exclusion
Execution time limit
An agent's single run is stopped when it goes past this time. Choose 5, 15, 30, 60, 90 or 120 minutes, or “No limit”. The default is “No limit”. A run that is taking a long time, or that is not responding, is also stopped once the limit is reached.
In Project, where the project's “Length of one run” is shorter, that is what stops it (Settings Reference).
DLP Integration
Asks your company's DLP (data loss prevention) endpoint about content before and after the agent uses tools (MCP, web fetch, file export). The execution platform calls the endpoint and does not send user or conversation IDs.
| Field | Content |
|---|---|
| Mode | “Off”, “Log only (check but do not stop)” or “Stop on violation”. The default is “Off” |
| Endpoint URL | An https URL. The endpoint returns {"decision": "allow"} or {"decision": "deny"} |
| Authentication | “Bearer token” or “mTLS (client certificate)” |
| Timeout | 100 to 25000 ms. The default is 3000 |
| Tools to check | Choose from “MCP”, “Web fetch” and “File export” |
- In “Stop on violation”, tool calls are stopped not only when the endpoint denies them but also when it cannot be reached, times out, or returns an unexpected format. Use “Test connection” first to make sure it is reachable
- When set to anything other than “Off”, the project step “External automation” cannot be created or run. The external API returns 409
automation_dlp_enabled
Prohibition Instructions in Common AI Agent Instructions
Under “AI Agents” > “Common AI Agent Instructions”, decide the text passed to the AI as instructions. Each setting has these items.
| Item | Details |
|---|---|
| User Prompt | Instructions that shape how the AI behaves (Markdown) |
| Prohibition Instructions | Instruction text for the AI, stating security policy and what is forbidden. Also used to check requests |
| Organization | “Global (all organizations)” is passed for every organization; choosing an organization passes it for that organization. The default is “Global (all organizations)” |
| Merge order (required) | 1, 2 or 3. The global settings are passed to the AI in this order, followed by the target organization's settings in this order |
| Status | Only active settings are passed |
- The target organization is the user's organization for the chat and scheduled runs, and the project's organization for project runs. When a project has no organization, only the global settings are passed
- Settings with the same merge order are all passed; their order among themselves is not fixed
- Deleting an active setting removes it from the instructions passed to the AI
- Requests judged to violate the Prohibition Instructions are not executed. Because the AI makes that judgment, anything you need stopped for certain, stop with Guardrails Settings or banned URLs. In project runs, execution rules can stop it too (What to decide on which screen)
When settings take effect
Changes to Guardrails Settings take effect within a few minutes, from the next request. The personal-information masking setting does not change partway through a request. The other checks (prompt attacks, harmful content, personal information) pick up a change within about 5 minutes, so in a long run they may change partway.
Prohibition Instructions in “Common AI Agent Instructions” can take up to 30 minutes to take effect.
Checking what was stopped
What guardrails detect is recorded under “Records and Audit” > “Guardrail Alerts”. The text itself is not kept, and nothing that identifies the user is shown on screen or in the API. Respond by informing users or adjusting settings, not by tracking down individuals (Logs and Audit).
Banned URL
Register destinations you do not want agents to reach as domain patterns (regular expressions). Example: .*example\.com
| Item | Details |
|---|---|
| Banned Domain Pattern (Regex) (required) | The pattern of domains to stop |
| Reason | Security, Compliance, Internal Policy, Other |
| Description | A note on its use or purpose |
| Active | Turned off, the pattern stays registered but stops nothing. Switching asks for confirmation |
- Registered destinations are always stopped. With nothing registered, agents can go out to any destination (Network)
- Use “Pattern Test” on screen to try URLs and confirm they are stopped before turning a pattern on
- Add a reason and a description so you can review later
- Do not register
localhostor private IP address ranges (such as10.*). The service would stop working properly
What was stopped or let through is recorded under “Records and Audit” > “Outbound Communication Records”.
Widget Connection Restrictions
An agent's answer can include a small interactive screen (a widget). By default a widget cannot connect out; it can connect only to the domains the agent declares.
Under “AI Agent Restrictions” > “Widget Connection Restrictions”, register the domains you do not want widgets to connect to as “Deny”. Deny takes priority over what the agent declares.
| Item | Details |
|---|---|
| Domain (required) | The host name to block. Every subdomain under it is blocked too. If you enter a URL, only the host name is saved |
| Note | Such as why you created it |
- Changes take effect within about 60 seconds
- Widgets that use a denied domain may not display correctly