rogue-security/rogue is a Python-based tool that checks whether AI agents follow work policies and behave as expected in defined scenarios, and it simulates adversarial attacks to find vulnerabilities. It attempts both normal and adversarial interactions together to detect policy violations and agent weaknesses. The repository calls itself an AI agent evaluator and red team platform, and the latest public release is v0.6.4.
The README splits Rogue into two parts. Automatic Evaluation validates agents based on scenarios and expected results, while Red Teaming looks for vulnerabilities using several attack techniques and also provides CVSS-based risk scores and compliance information. CVSS (Common Vulnerability Scoring System) is a standard framework for rating the severity of vulnerabilities. It can integrate with agents via A2A, MCP, and Python, and it also ships a CLI usable in a TUI and in CI/CD pipelines. The repository was created on June 5, 2025 and has 1,069 stars.
However, the changelog for v0.6.4 is not available in the given source, so what was added cannot be confirmed. The license is also only listed as NOASSERTION, so the terms of use are unverified. No figures comparing detection rates or false positive rates are presented either. Rogue is closer to an evaluation tool that reassesses current behavior against defined policies and scenarios and assigns risk grades. With only the public information available, it is difficult to judge the changes in this version or its detection performance.