Particle.news
Download on the App Store

Microsoft Open‑Sources RAMPART and Clarity to Put Safety into AI Development

The tools are meant to turn red‑team discoveries into repeatable tests and decision records that live with code so teams can treat AI safety as ongoing engineering.

Overview

  • Microsoft has published two open‑source projects named RAMPART and Clarity that are designed to embed safety work inside normal developer workflows rather than leave it to one‑off security reviews.
  • RAMPART is a pytest‑native test framework that lets engineers write adversarial and benign scenarios, plug an agent in with a thin adapter, run pass/fail checks in CI pipelines, and run statistical trials to catch probabilistic failures.
  • Clarity guides teams through problem definition, solution exploration and failure analysis and saves the results as .clarity‑protocol markdown files in the repository so decisions can be reviewed in pull requests and tracked over time.
  • The release responds to risks from agentic AI—such as prompt injection, unintended tool activation, data exfiltration and non‑deterministic regressions—by turning red‑team findings into automated checks that prevent repeat mistakes.
  • RAMPART builds on Microsoft’s earlier PyRIT automation for red‑teaming and the company says these tools will let engineers own safety checks, scale red‑team learnings into runnable assets, and make safety artifacts part of continuous development.