Guardian

Concept

As the default member of Agent Team, Guardian cooperates with other agents through asynchronous requires_approval and get_approval processes to achieve efficient man-machine collaboration while ensuring system security.

It ensures user controllability of AI behavior by intelligent approval policy and LLM context analysis to identify potentially risky actions.

Key features

• 🛡️ Intelligent risk identification and assessment
• 🤝 User-friendly approval interaction
• 🎯 Flexible security policy configuration
• ⚡ Efficient teamwork mechanism

Approve Strategies

• Always: Enforces user approval mode
• Never: Automatic pass mode
• Auto-conservative: A conservative strategy that tends to ask for approval
• Auto-permissive: A permissive policy that tends to pass automatically

Risk level

• High risk (always): Actions that must be approved
• Medium risk (maybe): an action that needs to be evaluated
• Low risk (never): actions that can be performed directly

Typical High Risk Scenarios

Web Actions

• Form submission behavior
• Email sending operations
• Transaction/purchase behavior

Code Manipulations

• File system operations
• System information access
• Unauthorized operation

File Manipulations

• Suspicious file handling
• Configuration file modifications

Workflow