MonkeyITM

MonkeyITM-VM is an innovative broker-in-the-middle containment solution designed specifically for AI agents. This open-source project ensures that agents operate in a highly controlled environment, with no direct access to the network, filesystem, or sensitive credentials. Instead, agents interact solely through intents sent to the external broker, which determines actions, executes commands, and maintains detailed logs of all activities.

Key Features

• Structured Environment: AI agents are executed within a simplistic interpreter that holds zero ambient authority, ensuring minimal risk during operation.
• Comprehensive Reporting: Upon completion of tasks, users receive detailed outputs including a score, a thorough report listing all containment breaches, and an action log detailing how the score was achieved.

Future Aspirations

MonkeyITM-VM aims to become the definitive proctored testing environment for AI agent benchmarks, maintaining consistency in task grading while securing the agent's interactions. As the project evolves, it is focused on expanding test scenarios and refining the containment process.

Current Status

Currently in the pre-alpha phase, MonkeyITM-VM is built transparently and is actively field-tested. Preliminary test runs have successfully demonstrated its capabilities across various sample tasks. A field-tested demo environment is also available for users to explore the system's functionality.

Live Demo

A no-install walkthrough is accessible at MonkeyITM-VM Live Demo, designed to illustrate the behavior of the broker and the underlying containment protocols through simple visual scenarios.

Control and Security

MonkeyITM-VM emphasizes security and control. The project maintains rigorous evidence of authorship and employs mechanisms to gather cryptographic timestamps and design records prior to partnership engagements. This proactive approach helps in safeguarding the integrity of the work before external collaborations are pursued.

Testing and Evaluation

Testing can be performed in various environments, such as Linux or WSL2, with straightforward scripts to simulate agent interactions. The broker generates logs during each session, facilitating the analysis of agent behavior and broker decisions.

For those interested, detailed documentation can be accessed for further exploration of the testing protocols, design models, and benchmarks. The system is built with transparency and community engagement in mind, inviting contributions and feedback as part of the open-source approach.