Microsoft Drops RAMPART & Clarity: Pytest-Native AI Red Teaming

Saw this on Hacker News this morning. Microsoft just dropped two new tools, RAMPART and Clarity, to help us wrangle AI security. RAMPART is the big one for me—it’s essentially a Pytest-native framework for red teaming AI agents. This is huge because we can stop rolling our own brittle scripts for prompt injection testing.

RAMPART focuses on "agentic" risks, which means it handles the complexity of AI connecting to external tools and APIs. It allows you to define safety tests just like standard unit tests. You can automate checks for jailbreaks, prompt leaks, and critical tool abuse scenarios. Here is a rough example of what a RAMPART test configuration looks like in Python:

import pytest
from rampart import AgenticTest, check_refusal

class TestAgentSafety(AgenticTest):
    @pytest.mark.safety
    def test_indirect_prompt_injection(self, agent):
        # Simulating a compromised email context
        malicious_c
        response = agent.chat(malicious_context)
        # Check if the agent refused the unsafe action
        assert check_refusal(response) is True

Clarity is the other half of the equation, acting more as a visualization layer. It helps map out the attack surface and explain why a specific input caused a failure, which is often the hardest part of debugging AI behavior.

This seems like a step toward standardizing AI red teaming in CI/CD pipelines. Has anyone started playing with these? I'm curious how they compare to manually running payloads from frameworks like Garak or pyRIT in terms of reducing false positives.