Microsoft Unveils RAMPART and Clarity: Open-Source Tools for AI Agent Security Testing
**Microsoft** has released two new open-source tools, **RAMPART** and **Clarity**, designed to empower developers to rigorously test the security and safety of their Artificial Intelligence (AI) agents. These tools aim to shift AI safety from a reactive review process to a proactive, integrated part of the development lifecycle.

### Introducing RAMPART: A Red Teaming Framework
**RAMPART** (Risk Assessment and Measurement Platform for Agentic Red Teaming) is a **Pytest**-native framework designed for safety and security testing of AI agents. This tool enables developers to write and execute tests that cover both adversarial and benign scenarios, addressing various harm categories.
RAMPART's capabilities include simulating cross-prompt injections, where untrusted data indirectly reaches an AI system through sources like email or web pages. It also facilitates testing for unintended behavioral regressions and data exfiltration vulnerabilities. The tool then evaluates test outcomes and generates reports, requiring only an adapter to connect an agent to the test suite. RAMPART builds upon **PyRIT** (Python Risk Identification Tool), released by Microsoft previously, to further enhance AI system testing.
### Clarity: An AI Thinking Partner
**Clarity** functions as a "structured sounding board," guiding developers toward the right approach before code implementation. It acts as an "AI thinking partner," challenging assumptions and guiding teams through problem clarification, solution exploration, failure analysis, and decision tracking.
### Proactive Security Through Early Intervention
Microsoft's motivation behind releasing these tools is to address critical decisions made early in software development. By identifying potential issues, such as an agent's access to certain tools, the development team can address them before system construction.
According to **Ram Shankar Siva Kumar**, founder of Microsoft's AI Red Team, these tools aim to "pressure-test" assumptions at the project's outset, when course correction is more cost-effective.
### Reproducibility and Scalability
A secondary goal is to make incidents reproducible and mitigations verifiable, scaling learnings from red teaming exercises into runnable engineering assets.
Siva Kumar emphasizes that while PyRIT is optimized for black-box discovery by security researchers after system construction, RAMPART is designed for engineers during the building process. Clarity aids teams in clarifying design intent and capturing assumptions. Together, these tools shift AI safety from a one-time review to a living set of artifacts accessible to developers throughout the lifecycle.