Permission Management & Safe Usage

MBAgent is an AI that truly can take action — it operates your files, executes commands, and operates Mbbrowser accounts. Precisely because of this, permission management and safe usage are critical.

1. MBAgent Permission Layers

MBAgent's permission system is divided into 5 layers:

Layer 5:Mbbrowser Client ControlV2 protocol-level restriction← Mbbrowser's own security policy

Layer 4:User permission mode (Ask / Auto / Plan)← MBAgent Settings

Layer 3:Tool approval (each tool can be set individually)← Tool-level approval

Layer 2:File sandbox (limited to working directory)← Workspace isolation

Layer 1:Operating system permissions← Your computer's account permissions

2. Permission Modes: Ask / Auto / Plan

1. Ask Mode (Default Recommended)

Each high-risk operation pops up for your confirmation.

Target audience:

✅ Newbie users
✅ Scenarios involving important data / accounts
✅ When unsure what AI will do

Operation flow:

AI: I want to save the file to ~/reports.xlsx
MBAgent: popup
  [Title] Write file
  [Content] AI wants to write to ~/reports.xlsx. Allow?
  [Buttons] Allow / Deny / Allow and remember

2. Auto Mode

Common operations auto-execute, only very few popups.

Target audience:

✅ Advanced users familiar with MBAgent
✅ Repetitive tasks (let AI run itself to completion)
✅ Sandbox environment (no fear of AI breaking things)

Auto-executed operations:

Read files
Search web
Call read-only API

Still popups operations:

Write files
Execute shell commands
Operate Mbbrowser accounts

3. Plan Mode

AI first submits a plan, each step needs your approval.

Target audience:

✅ Complex multi-step tasks
✅ Want to understand AI's thinking
✅ Don't want to be interrupted by frequent popups

Operation flow:

AI: I plan to execute the following plan:
  1. Open 10 store environments
  2. Collect order data
  3. Write to Excel
  4. Close environments
  
Please approve:
  [Approve] [Cancel] [Modify Plan]

Switch Mode

Go to "Settings → Permissions"
Select mode
Save

3. Tool Approval

Each AI tool can have its own approval policy.

Default Approval Policy

Tool	Ask Mode	Auto Mode	Plan Mode
`Read` (read file)	Auto	Auto	Planning
`Write` (write file)	Popup	Popup	Planning
`Edit` (edit file)	Popup	Popup	Planning
`Bash` (execute command)	Popup	Popup	Planning
`WebFetch` (fetch page)	Auto	Auto	Planning
`WebSearch` (search)	Auto	Auto	Planning
Control Mbbrowser environment	Popup	Popup	Planning
Form filling / submit	Popup	Popup	Planning

Custom Approval Policy

Go to "Settings → Permissions → Tool Approval"
Find the tool to adjust
Select: Auto / Popup / Forbidden

4. File Sandbox

Working Directory Limit

All file operations in MBAgent can only be within the working directory you specified:

✓ /my-project/        ← Working directory, can read and write
✗ C:\Windows\        ← System directory, access forbidden
✗ D:\Private\        ← Outside working directory, access forbidden

Modify Working Directory

Click the "Working Directory" button on MBAgent main UI
Select new directory
MBAgent reinitializes the sandbox

IMPORTANT

Confirm working directory at each startup: avoid MBAgent mistakenly operating a directory you didn't intend to let it touch.

Files Outside Sandbox

When you need to access files outside the working directory:

Temporary: Copy files into the working directory
Permanent: Modify working directory to include the target location
Dangerous: Turn off sandbox in settings (not recommended)

5. Command Execution Limits

Default Blocked Dangerous Commands

MBAgent has a built-in dangerous command blacklist that blocks by default:

Category	Blocked commands
System-level	`format`, `diskpart`, `bcdedit`, `reg delete`
Network attacks	`netsh firewall`, `net user /add`
Deletion	`rm -rf /`, `del /f /s /q C:\`
Process operations	`taskkill /f /im`, `ps -ef \| grep` combined with kill
Download exec	`curl ... \| bash`, `wget ... \| sh`

WARNING

Blacklist is a fallback, not absolute protection. Complex attack commands may bypass. It is recommended to still use Ask mode in important environments.

Custom Dangerous Commands

Go to "Settings → Permissions → Command Blacklist"
Add command patterns you want to block
Save

6. API Key & Credential Management

Storage Location

Credential Type	Storage Location	Encrypted
Mbbrowser Account Token	Mbbrowser local database	✅
AI Provider API Key	Windows Credential Manager	✅
Third-party account password	Temporary in current session only	❌ (memory)
Mbbrowser account password	Used at login only, not stored	❌ (memory)

Credential Usage Flow

[1] User logs into Mbbrowser
    Mbbrowser account + password → Mbbrowser client validates
    ↓
[2] Mbbrowser client returns Token
    Token written to local database
    ↓
[3] MBAgent startup
    Reads Token from Mbbrowser client
    ↓
[4] MBAgent uses Token to connect ControlV2
    Token never passes through frontend chat content

Credential Security Recommendations

✅ Regular rotation: Periodically change password in Mbbrowser console
✅ Two-factor verification: Enable 2FA on Mbbrowser account
✅ Independent accounts: AI operations use independent Mbbrowser sub-accounts
✅ Revocation mechanism: Immediately revoke Token in Mbbrowser console when anomalies are found

7. AI Memory Bank Security

Memory Bank File Visibility

AI memory bank files are stored in:

%LOCALAPPDATA%\mbagent\memory\

Plain text — anyone who can access your computer can read it.

What Not to Store in Memory Bank

❌ Never store:

Passwords
Tokens / API Keys
Bank card numbers
ID numbers

✅ Can store:

Your preferences
Account numbers (not passwords)
Business rules
Public facts

Encrypted Backup

If you need to back up sensitive memories, recommended:

Use 7-Zip encrypted compression
Store on encrypted mobile HDD
Don't upload to cloud drives

8. Security Audit & Logs

Operation Logs

MBAgent records all AI operations:

[2026-09-24 22:30:15] AI calls tool Write
  Params: {"file_path": "D:/my-project/reports.xlsx"}
  Result: Success
  Token: 1234

[2026-09-24 22:30:18] AI calls Mbbrowser ControlV2
  Params: {"action": "browser.start", "env_id": "223223"}
  Result: Success
  Duration: 3.2s

View Logs

Go to "Settings → Run Logs"
Enable "MBAgent Log"
Log file location:
```
%LOCALAPPDATA%\mbagent\logs\
```

Anomaly Detection

If suspicious operation is found:

Immediately click "Stop Current Task" in MBAgent
Revoke that environment's Token inside Mbbrowser
Modify Mbbrowser account password
Contact Mbbrowser customer support

9. Best Practices

For Individual Users

✅ Use Ask mode (default)
✅ Confirm before important operations
✅ Periodically check AI memory bank
✅ Don't let AI operate banking / payment sites

For Enterprise Users

✅ Use independent Mbbrowser sub-accounts for AI
✅ Limit AI's accessible environment range
✅ Enable Mbbrowser operation audit logs
✅ Periodically rotate AI's API Keys
✅ Provide MBAgent security training to employees

Dangerous Operation List

⚠️ What MBAgent should NOT do (even if it claims it can):

❌ Modify system files / registry
❌ Install / uninstall software
❌ Delete databases
❌ Operate payment / banking sites
❌ Send external emails (unless explicitly allowed)
❌ Publish content on social platforms

10. Emergency Handling

Immediately Stop AI

Method 1: Press Esc key
Method 2: Click the "Stop" button in the input box
Method 3: Close the MBAgent main window
Method 4: Close all environments inside Mbbrowser

AI Performed an Incorrect Operation

Immediately stop (see above)
Check AI memory bank, delete wrong memories
Check operation logs, see what AI did
Manually revoke AI's incorrect operation
Add "operations that shouldn't be done" to the command blacklist

Suspect Account Stolen

Immediately change password in Mbbrowser console
Revoke all Tokens
Check Mbbrowser account operation logs
Contact Mbbrowser customer support

Next: Start Menu Tasks → Save frequently-used tasks as buttons.

Permission Management & Safe Usage ​

1. MBAgent Permission Layers ​

2. Permission Modes: Ask / Auto / Plan ​

1. Ask Mode (Default Recommended) ​

2. Auto Mode ​

3. Plan Mode ​

Switch Mode ​

3. Tool Approval ​

Default Approval Policy ​

Custom Approval Policy ​

4. File Sandbox ​

Working Directory Limit ​

Modify Working Directory ​

Files Outside Sandbox ​

5. Command Execution Limits ​

Default Blocked Dangerous Commands ​

Custom Dangerous Commands ​

6. API Key & Credential Management ​

Storage Location ​

Credential Usage Flow ​

Credential Security Recommendations ​

7. AI Memory Bank Security ​

Memory Bank File Visibility ​

What Not to Store in Memory Bank ​

Encrypted Backup ​

8. Security Audit & Logs ​

Operation Logs ​

View Logs ​

Anomaly Detection ​

9. Best Practices ​

For Individual Users ​

For Enterprise Users ​

Dangerous Operation List ​

10. Emergency Handling ​

Immediately Stop AI ​

AI Performed an Incorrect Operation ​

Suspect Account Stolen ​

Permission Management & Safe Usage

1. MBAgent Permission Layers

2. Permission Modes: Ask / Auto / Plan

1. Ask Mode (Default Recommended)

2. Auto Mode

3. Plan Mode

Switch Mode

3. Tool Approval

Default Approval Policy

Custom Approval Policy

4. File Sandbox

Working Directory Limit

Modify Working Directory

Files Outside Sandbox

5. Command Execution Limits

Default Blocked Dangerous Commands

Custom Dangerous Commands

6. API Key & Credential Management

Storage Location

Credential Usage Flow

Credential Security Recommendations

7. AI Memory Bank Security

Memory Bank File Visibility

What Not to Store in Memory Bank

Encrypted Backup

8. Security Audit & Logs

Operation Logs

View Logs

Anomaly Detection

9. Best Practices

For Individual Users

For Enterprise Users

Dangerous Operation List

10. Emergency Handling

Immediately Stop AI

AI Performed an Incorrect Operation

Suspect Account Stolen