RAMPART is an open-source safety and security testing framework for agentic AI applications, built by Microsoft and published on GitHub under the MIT license. Its name stands for Risk Assessment and Measurement Platform for Agentic Red Teaming, and with more than 420 stars and updates in October 2026, it is actively maintained.
RAMPART gives developers a structured, developer-friendly way to write and run safety tests for AI agents. It covers adversarial attacks, benign failure modes, and a broad range of harm categories, expressing each check as an evaluation-driven assertion. Because it is pytest-native, the tests live next to ordinary unit tests, run in the same CI pipelines, and report failures the way Python teams already expect. The project also ships an onboarding skill for coding assistants that helps a team decide what a first test should prove, build the integration, run it, and then choose code-grounded follow-up tests and payloads.
The repository carries an OpenSSF Scorecard badge and continuous integration through GitHub Actions, and releases are published to PyPI for current Python versions. For teams, that means RAMPART fits the same release discipline as the rest of a Python codebase: versioned, reviewable tests that fail a build when an agent regresses on a known attack pattern or harm scenario, instead of problems being rediscovered by hand before every launch.
RAMPART is free and open source under the MIT license. There is no paid tier or hosted service; you install the package from PyPI and run the tests inside your own environment and CI.
RAMPART suits engineering and AI safety teams shipping agentic applications who want repeatable red-teaming and regression checks for agent behavior, rather than one-off manual prompt testing before a release.