Black-Box Red Teaming Framework for Agentic AI Risk Discovery

Summary

Agentic systems increasingly process untrusted inputs, use tools with real permissions, and act across multiple steps, creating risks that single-turn evaluations can miss. This study presents a systematic black-box framework that requires only basic system descriptions, allowing evaluators to test agents without privileged access. The framework organizes observable behavior into seven risk domains and uses an automated system called SAGE-RT to generate 120 adversarial scenarios for each domain. Human-validated evaluation is performed with the help of LLM judges. The authors test the approach on two agent architectures, CrewAI and AutoGen, using four base models. Across the evaluation, average governance risk reaches 56.25%, privacy risk reaches 65% in multi-agent configurations, and agent behavior vulnerabilities reach 85%. The results indicate that important weaknesses can arise from the architecture and multi-step operation of agents, rather than from chat responses alone. The authors present the method as a scalable way to identify critical vulnerabilities during safer agent deployment, while the abstract does not provide further details about the tested models, scenarios, or mitigation performance.