Project Chaperone: A Rule-Enforcing Harness

(appliedingenuity.substack.com)

5 points | by funeric555 4 hours ago ago

2 comments

  • funeric555 3 hours ago ago

    From the post: Announcing Project Chaperone:

    AI agents have been known to delete databases, leak data, or make up legal documents.

    The usual fix is a line in the prompt: "NEVER send personal data out in e-mail." A prompt is a request. Nothing enforces it.

    Dan DeMaggio and I built Project Chaperone, an open-source harness that does two things:

       1: Before it runs, it reports every place a workflow could break your rules.
       2: While it is running, it blocks any call that would break a rule.
    
    It works using labels on data. Data from your inbox is labeled PERSONAL. Labels stay on the data, even after an LLM rewrites it. If a rule says no PERSONAL data goes out by e-mail, and an agent tries to call an e-mail send module with personal data, the harness blocks that call.

    As a first demo project (in the released code) we built an AI-based system that can read e-mail, locate delivery messages, extract the info and update a calendar. I trust the harness so much that I run it on my real e-mail and calendar. The worst it can do is add a wrong calendar entry.

    This harness does not make an LLM perfect. It makes the rules explicit, and it enforces them.

    Paper: https://appliedingenuity.substack.com/p/project-chaperone-a-... and code (MIT license) https://github.com/AppliedIngenuity-ai/project-chaperone

  • funeric555 4 hours ago ago

    [flagged]