autogen - 💡(How to fix) Fix Proposal: Standardized Safety Sandbox for Agent Tool Execution [3 comments, 3 participants]

Fix Plan

To address the issue, we'll introduce a ToolSafetyPolicy interface and implement it for different executors.

Step-by-Step Solution:

• Define the ToolSafetyPolicy interface with methods for setting safety constraints:
  • allow_network
  • allow_filesystem
  • max_memory
• Implement the interface for each executor (e.g., Docker, Firecracker, WASM)
• Modify the tool definition to include safety constraints
• Update the executor to enforce the safety policy before running the tool

Example Code (Python):

from abc import ABC, abstractmethod

class ToolSafetyPolicy(ABC):
    @abstractmethod
    def allow_network(self, allow: bool):
        pass

    @abstractmethod
    def allow_filesystem(self, allow: bool):
        pass

    @abstractmethod
    def set_max_memory(self, max_memory: int):
        pass

class DockerExecutor(ToolSafetyPolicy):
    def allow_network(self, allow: bool):
        # Implement Docker-specific network allowance
        print(f"Allow network: {allow}")

    def allow_filesystem(self, allow: bool):
        # Implement Docker-specific filesystem allowance
        print(f"Allow filesystem: {allow}")

    def set_max_memory(self, max_memory: int):
        # Implement Docker-specific memory limitation
        print(f"Set max memory: {max_memory} MB")

# Example usage:
docker_executor = DockerExecutor()
docker_executor.allow_network(True)
docker_executor.allow_filesystem(False)
docker_executor.set_max_memory(1024)

Verification

To verify the fix, test the ToolSafetyPolicy implementation for each executor and ensure that the safety constraints are enforced correctly.

Extra Tips

• Consider using a configuration file or database to store the safety constraints for each tool and executor.
• Implement logging and monitoring to detect and respond to potential security incidents.
• Regularly review and update the ToolSafetyPolicy interface and implementations to ensure they align with emerging safety standards in AI.