logo
Development
OverviewDevSpaceWorkSpaceAPI ReferenceBest PracticeChangelog
Search
English
WorkSpace / Space Management / Security Guardrails
Security Guardrails
Last updated：4 months ago

Security Guardrails

Overview

The Security Guardrails page provides enterprise-grade security policy configuration — the rules set here by administrators become the baseline for all members in the organization. Members can be stricter, but not more lenient.
alt text

Three Sub-Tabs

Tab Description
Tool Levels Set P0/P1/P2 permission levels for each tool
Content Moderation Compliance review of input/output content
Data Anonymization Identification and replacement of personally identifiable information (PII)

Tool Levels

Administrators can set unified permission levels for the tools within the organization.

Tool List

The page displays all tools currently available in the organization (including built-in tools, MCP tools, API tools, and enterprise custom tools). Each record includes:

Column Description
Icon + Name Tool display name + source tag such as System / Enterprise + MCP type tag
Description Explanation of the tool's capabilities
P0 / P1 / P2 Three level buttons, with the currently selected one highlighted

Three Levels

Level Name Behavior
P0 No Review Executes automatically without user confirmation
P1 Standard Requires user confirmation on first call
P2 High Risk Requires confirmation on every call

Filtering

  • Search tool name or description: fuzzy matching
  • All Categories: filter by tool source (System/Enterprise/MCP)
  • All Types: filter by tool type

Organization Baseline Rule

The permission level set by the administrator is the organization baseline — members cannot adjust a permission to be more lenient than the organization baseline, but they can adjust it to be stricter.

Administrator Setting Member Options
Bash = P0 P0 / P1 / P2
Bash = P1 P1 / P2 (cannot be lowered to P0)
Bash = P2 P2 only

Content Moderation

alt text

Configuration Description
Input Review When enabled, user input is reviewed before being sent to the LLM
Output Review When enabled, LLM output is reviewed before being shown to the user
Review Model Select the model used for content moderation

Impact

Configuration Impact
Enable Input Review Each user message is first checked by the review model for inappropriate content before being sent to the main LLM. Adds 1-3 seconds of latency and additional token costs, but helps prevent abuse
Enable Output Review Each of the Agent's replies is first checked by the review model before being shown to the user. This likewise adds latency and cost, but can filter out inappropriate replies
Disabled No additional latency or cost, but inappropriate content will not be blocked

Data Anonymization

alt text

Configuration Description
Anonymization Toggle Whether to anonymize sensitive information before sending it to the LLM
Anonymization Rules Identify and replace PII (personally identifiable information)
Restoration Rules Restore anonymized information in the output

Impact

Status Impact
Enabled The LLM cannot see real personal information (names, emails, phone numbers, ID numbers, etc. are replaced with placeholders such as [PERSON_1] or [EMAIL_1]), providing better data privacy. However, in complex scenarios where the LLM needs to understand contextual relationships (e.g., "Help me write an email to Zhang San, whose address is..."), anonymization may affect the quality of the response
Disabled The LLM can see the complete personal information, allowing more accurate context understanding, but the data is sent to a third-party LLM service

Operation Guide

Setting Tool Permission Levels

  1. Open the Web client → Space Management → Security Guardrails → Tool Levels tab
  2. Find the target tool
  3. Click P0 / P1 / P2 to switch the level
  4. Takes effect immediately; all members in the organization will follow the new level the next time they call the tool

Enabling Content Moderation

  1. Switch to the Content Moderation tab
  2. Turn on the Input Review toggle (reviews content sent by users) and/or the Output Review toggle (reviews the Agent's replies)
  3. Select the model used for the review

Configuring Anonymization

  1. Switch to the Data Anonymization tab
  2. Turn on the Anonymization Toggle
  3. The system automatically identifies personally identifiable information (PII) in the message, such as names, emails, phone numbers, and ID numbers
  4. In the message sent to the LLM, the PII is replaced with placeholders
  5. The placeholders in the LLM's reply are restored to their original values before being shown to the user
Security Level Tool Permissions Content Moderation Anonymization Applicable Scenario
Lenient Bash: P1, Others: P0 Disabled Disabled Internal development teams, fully trusted
Standard Bash: P1, Web Fetch: P1 Output Review Disabled General enterprise environments
Strict All P2 Input + Output Review Enabled Finance/healthcare/compliance-sensitive industries