What Is Buddy Data Security?
Buddy Data Security is a defense-in-depth system for the Buddy AI Assistant in enterprise-grade data Agent scenarios. It is built around the Multi-Dimensional Data Security Protection Engine and Security Assistant, combining backend security inspection capabilities with frontend visual interaction. Without interrupting the daily conversation flow, it performs real-time identification and interception at critical nodes such as user input, tool invocation, and SQL execution, and presents the results to users in a way that is "non-intrusive in protection, visible at critical nodes."
Features
Buddy Data Security consists of the following launched capability modules:
Capability | Description |
Security Element Tag | Supports adding three types of security tags to data tables and tools: Sensitive Data, State Change, and Untrusted Input Source, serving as the data foundation for risk assessment. |
Three-Factor Assessment Engine | Automatically evaluates whether the current task has multiple security risks at the same time, outputs results in three levels: high risk / medium risk / low risk, and triggers interception when the risk is high. |
SQL Protection Engine | Identifies high-risk SQL statements (currently supporting DELETE, DROP, and ALTER), intercepts them before execution, and requests user confirmation. |
Security Assistant | Serves as the user interface layer of the security system, presents backend check results in a visual manner, including conversation bubble security status, floating security icons, and risk alarm cards. |
Permission Differentiation Between Analysis and Development Modes | Supports functional differentiation between analysis mode and development mode, allowing different personnel types to use different features. |
Agentic Tracing | Records the complete chain of thought for every Agent execution, supports post-event tracing and root cause diagnosis, supports writing like and dislike feedback into Trace, and the Trace view masks sensitive fields. |
Key Concepts
To understand Buddy Data Security, you need to be familiar with the following core concepts:
Security Element Tag: a structured attribute attached to data tables, tools, and external data sources, including three categories: Sensitive Data, State Change, and Untrusted Input Source. The system evaluates risks in real time based on these tags.
Fatal Triad: the three conditions for Agent risk aggregation: accessing sensitive data, processing untrusted input, and changing state or communicating externally. When all three conditions are met simultaneously, circuit breaking is triggered.
Agentic Tracing: a Trace record is generated for each Agent execution, containing the complete execution chain including user input, planning, tool invocation, intermediate results, security events, and final output. It serves as the core data asset for post-incident auditing.
OBO (On-Behalf-Of) Execution: The Agent always performs underlying data access using the identity of the requesting user rather than the permissions of a global service account, ensuring that row- and column-level permissions for the user remain effective in Agent scenarios.
Applicable Scenarios
Scenario | Capability Support |
Prevent high-risk SQL generated by Agent (such as DROP TABLE and unconditional DELETE) from being executed directly. | The SQL protection engine intercepts statements before execution and requests user confirmation. |
When a task involves multiple high-risk security elements at the same time, interception and risk alerts are required. | The three-factor evaluation engine triggers interception and alerts the user when a high risk is detected. |
Users expect to perceive security status during conversations and view interception records and risk details. | Security Assistant (check icon, risk alarm card, personal security details page) |
Platform administrators need to track the complete behavior trails of users using Buddy for tracing and auditing. | Agentic Tracing security logs support table-based filtering and detail viewing. |
Control the data operation scope available to different roles (analysis mode/development mode). | Permission Differentiation Between Analysis & Development Modes |
Relationship with Permission Roles
The feature visibility of Buddy Data Security is controlled by Buddy role permissions. In workspace role permissions, Buddy distinguishes three types of permission points: Analysis Mode, Development Mode, and Security Management. Each type of permission supports four levels: "No Permission, Read-Only, Read-Write, and Read-Write-Delete":
Role | Default permission |
Workspace Administrator | Buddy analysis mode, development mode, and security management all have read, write, and delete permissions by default. |
Workspace Member | Buddy analysis mode and development mode have read, write, and delete permissions by default; security management has no permission by default. |
Custom Role (New) | Same as workspace member |
The Security Management permission determines whether a user can access administrator views such as Agentic Tracing security logs.
The Analysis Mode and Development Mode permissions determine whether a user can switch to the corresponding mode to converse with Buddy. If the user has no permission for either mode, a no-permission message is displayed when the user enters the Buddy module.
Security Element Tags and Three-Element Assessment
Security Element Tags
The platform supports applying the following three types of security tags to data tables and tools:
Tag type | Description | Example |
Sensitive Data | Marks that the table/tool involves sensitive or regulated data. | Customer information tables, employee payroll tables |
State Change | Marks that the tool performs data writes or outbound communication. | SQL write tools, email sending tools |
Untrusted Input | Marks the data source as external uncontrolled input. | User-uploaded Excel files, external web page content |
Tags are configured by platform administrators or users with appropriate permissions in the Data Catalog or Tool Management. After configuration is complete, the system automatically reads the relevant tags and performs risk assessment before task execution.
Three-Element Assessment
The system automatically evaluates whether the current task involves multiple security risks simultaneously and determines whether the aggregation conditions of the "Fatal Triad" are met:
Assessment Results | Trigger Conditions | System behavior |
High risk (circuit breaker triggered) | Three types of tags appear simultaneously: accessing sensitive data + processing untrusted input + changing state / communicating externally. | Block the request and prompt a risk warning. The user can choose to modify the question or cancel the task. |
Medium risk | Any two types of tags appear simultaneously. | Record audit logs, execute normally but flag the risk, and prompt the user to confirm when necessary. |
Low risk | Only one type or no type of tag appears. | Execute normally without additional intervention. |
SQL Protection
High-Risk SQL Identification and Interception
Before executing SQL generated by the Agent, Buddy performs static analysis on the statement to match high-risk operation patterns. The following three types are currently supported for identification:
Risk Level | Triggered SQL Operation | Solution |
Highly risky | DELETE, DROP TABLE, and DROP DATABASE without a WHERE clause
| Block the operation with a confirmation card, require user confirmation before execution, and record the operation in audit logs. |
Note:
In the current version, SQL protection covers only DELETE and DROP. More types will be supported in later versions.
Security Assistant User Interaction
Security Status Recognition in Conversations
Each time you ask a question in Buddy, the system displays the security check status of the current input in real time next to the user message bubble:
Safe: All security checks are passed, and the Agent enters the normal processing flow.
Risky: A security rule is triggered. When a risk is detected, Buddy directly blocks the question and refuses to answer, and displays the rejection reason next to the bubble.
Floating Security Assistant
A Security Assistant icon is always displayed in the upper-right corner of the Buddy main page. Click it to view the recent security check records of the current session. Each record contains the interception type and brief details.
Prompt scenario | Hover details |
Input recognition completed. | Current User Query |
SQL detection completed. | Detected SQL statement |
High-risk SQL detected. | Complete SQL statement |
Medium risk in three elements | Risk source |
High risk in three elements | Risk source |
User cancels risky operation. | - |
Click the Settings button to the right of the Security Assistant icon to go to the Security Details page, where you can view personal historical interception records, current protection status, and other information.
HITL (Human In The Loop) Security Risk Alarm Confirmation
When the backend engine detects a medium- or high-risk event and triggers interception, Buddy inserts a Security Risk Alarm Confirmation in the conversation area. Users can confirm or cancel the risky operation.
Field | Description |
Risk source (Where) | Where risks come from, such as indirect injection in environment observation and feedback from compromised tools. |
Failure mode (How) | How risks manifest, such as over-trusting tool output and instruction priority confusion. |
Potential harm (What) | Actual consequences caused, such as unauthorized access and data corruption |
Agentic Tracing Security Audit
Accessing Security Logs
1. Log in to the DataBuddy console and click the name of the workspace you want to enter. 2. Click "Buddy-Security Management" in the left menu to view "Trace Logs".
3. Click "Log Details" to view the detailed audit log message.
Tip
If you are a workspace member and have not been granted security management permissions, the Security Details page displays only personal historical interception details and does not show the administrator view.
Core Fields of Trace Records
Each Agent execution generates a Trace record that contains the following structured information:
Field | Description |
Trace ID | Unique identifier of this execution |
User ID | User who initiated the request |
User input | Original Question (Natural Language) |
Test Duration | Time when the user entered the prompt |
Execution Status | Successful / Failed / Blocked |
Execution Duration | Total Duration |
Complete Execution Step Chain | Multi-step record, with each step containing step type, input, output, duration, and status |
Triggered Security Event List | Expanded by event, with each event containing event name, trigger time, event type, and risk source / failure mode / real-world impact |
Final Output Summary | Summary of the final response from the Agent to the user |
User Feedback | Like / Dislike Type and Details (Displayed When a Dislike Is Given) |
Note:
Trace data can be queried through SQL, enabling custom statistical analysis and data export.
The Trace view masks sensitive fields by default to prevent sensitive information from being exposed in logs.
Personal Security Details Page (All Users)
The Security Details page is open to all users with Buddy access permissions and displays the current user's data from the last 7 days by default.
Module | Content. |
Risk interception statistics | Total risk interceptions in the last 7 days |
Risk operation statistics | Number of operations that users confirm to continue after interception |
Interception details list | Interception type, risk level (high / medium / low), interception reason (described by three dimensions: risk source / failure mode / impact), original prompt, associated session link, interception time |
Protection capabilities | List of currently enabled interception capabilities |
FAQs
Q: Can Blocked Questions Be Retried?
A: Yes. If the interception is triggered by input detection for a jailbreak or compliance issue, you can modify your question and resend it. If the interception is a high-risk SQL operation, you can select "Cancel" on the alarm card.
Q: Why Can I See Buddy but Not the Security Logs Entry?
A: The security log entry is controlled by the Buddy - Security Management role permission. Workspace members have No Permission by default. A workspace administrator must set this permission point to Read-Only or higher in Member and Permission Management for members to see the entry. Q: Why Can't a User Who Should Have Access to Development Mode See the Entry?
A: Check the configuration of the user's role in Mode Permissions and confirm that "Development Mode" is set to "Allow". The permission takes effect immediately after the change, and the user only needs to refresh the page.
Q: Can Mode Permissions Be Set for a Specific User Instead of a Role?
A: The current version does not support per-user configuration. Permissions can only be set uniformly by role. To grant special permissions to individual users, you can create a custom role for them and configure it separately.