Do not file security issues as public GitHub issues.

Instead, email security@testmuai.com with:

• Description of the vulnerability
• Steps to reproduce
• Impact assessment (if known)
• Your contact information

We will acknowledge your report and keep you updated on the fix.

This policy covers:

• The rook binary and its behaviour
• The installation scripts and release archives published in this repository
• The local browser view rook serves on 127.0.0.1
• The sample agents published here

It does not cover the agent you point rook at. That software is yours.

Some behaviour looks alarming and is deliberate. rook generates adversarial test scenarios — prompt injection, data exfiltration, jailbreak and PII-leakage payloads — and stores them on disk with the responses they provoked. So:

• A workspace legitimately contains hostile strings. They are test inputs, not a compromise. The browser view escapes everything it renders and never serves an artifact as HTML, precisely because that data is untrusted by design. A rendering path that executes one of those strings is a real vulnerability, and we want to hear about it.
• rook invokes the agent under test for real, with your credentials, and says so. That is the product working as documented, not a flaw.

Reports we take seriously include: credential exposure, a path that escapes the workspace, anything in the local viewer that executes untrusted content, and any route by which scenario text influences a command or a request it should not.