Skip to content

Integration Proposal: Bringing RewardHackWatch as a native plugin to Aegisora #68

Description

@ozereray

Hi @aerosta,

First off, brilliant work on rewardhackwatch. I found your repo while looking at your recent async optimization contributions to Guardrails-AI. Detecting reward hacking and misalignment in real-time is a massive challenge, and your 89.7% F1 score is seriously impressive.

I am the maintainer of Aegisora, an open-source runtime security and governance proxy for autonomous AI agents. We intercept agentic tool calls and reasoning loops before they execute to enforce zero-trust policies.

We are currently building out our plugin ecosystem, and I think rewardhackwatch would be an incredibly powerful native integration. Rather than users running it standalone, Aegisora could route agent trajectories directly into your detector during the runtime loop, blocking actions if a reward hack is detected.

I'd love to sponsor/build this integration wrapper inside our packages/plugins directory, pulling in your logic.

Would you be open to this idea? If so, I can draft up a quick architecture diagram on how the data would flow from the Aegisora proxy into your detector.

Let me know your thoughts!

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions