TrinityGuard
A Python framework for evaluating safety risks in multi-agent systems, wrapping an existing MAS to run structured checks across 20 risk types such as prompt injection, tool misuse, and sandbox escape, and producing evidence-backed safety reports.
TrinityGuard wraps a multi-agent system through its main entry point to run a task, trace what happens, and generate a structured safety report, independent of the specific agent framework underneath. It evaluates behavior across 20 built-in risk categories organized into three severity levels, including prompt injection, sensitive-data disclosure, tool misuse, message tampering, cascading failures, sandbox escape, and rogue-agent behavior, and supports LLM-as-judge scoring alongside monitor observations and calibration data.
It ships integration paths for two popular multi-agent frameworks, and an experimental adapter for another agent runtime, alongside a framework-independent execution layer for tracing workflows and intercepting messages. An opt-in runtime-protection mode can evaluate live messages and return allow, replace, or deny decisions rather than only reporting after the fact.
It requires Python 3.10 or newer and installs from source; example runs include a deterministic in-process multi-agent system for checking the public API before wiring in a real framework adapter.