MonkeyITM – open-source broker that sits between agents and the open world (github.com)

MonkeyITM has announced an open-source broker designed to enhance the containment of AI agents, enabling a secure interaction layer between agents and their external environments. Operating within a tightly controlled interpreter, these agents are stripped of network, filesystem, or credential access, relying solely on a broker that logs intents, executes actions, and assesses agent performance. This architecture aims to create a proctored exam environment for AI benchmarks, producing standardized scores and detailed containment reports that track any attempts by agents to exceed their intended capabilities. Significantly, MonkeyITM offers a rigorous framework for evaluating AI behaviors under controlled conditions, greatly enhancing the transparency and safety of AI experiments. Currently in pre-alpha, the project is built using Rust and Python 3, featuring a simulated CLI test environment where real models can be assessed for compliance with containment protocols. The initial testing phase has already demonstrated successful containment of known models, paving the way for future scenarios with hidden tests to assess the agents further. Overall, MonkeyITM's approach not only reinforces security in AI operations but also establishes a foundation for responsible AI benchmarking.