RIFT-Bench Enables Dynamic Red-Teaming for Agentic AI

Agentic AI systems, powered by large language models, are rapidly evolving into autonomous decision-making entities, introducing new attack vectors beyond those found in traditional LLM vulnerabilities. Existing security evaluation methods are often tied to specific implementations or domains, making unified comparisons across heterogeneous systems challenging. To address this gap, RIFT-Bench has been developed as a graph representation-driven methodology for dynamic red-teaming. This framework enables unified security evaluations across a wide range of agentic architectures. It operates in two automated phases: a Discovery phase that extracts the system's internal structure, and a Scanning phase that deploys adaptive adversarial attacks. RIFT-Bench evaluates the system itself by leveraging a broad set of dynamically adaptable adversarial probes across diverse attack vectors and objectives. The effectiveness of this evaluation pipeline was demonstrated across 45 agentic systems, showcasing its ability to generalize to heterogeneous architectures. Beyond assessing systems and attacks, RIFT-Bench also supports the direct evaluation of mitigation strategies, establishing itself as a scalable foundation for the security assessment of agentic AI systems.