Researchers propose RIFT-Bench, a dynamic red-teaming framework for evaluating agentic AI systems

Researchers from an unnamed set of contributors introduced RIFT-Bench, a graph representation–driven methodology for dynamic red-teaming designed to evaluate agentic AI systems. The framework is motivated by the observation that agentic systems—powered by large language models (LLMs)—are evolving into autonomous decision-making entities with attack vectors that extend beyond those of traditional LLM vulnerabilities.

RIFT-Bench operationalizes evaluation through two automated phases: Discovery and Scanning. In the Discovery phase, the system extracts the structure of the target agentic architecture using a hierarchical graph representation. The Scanning phase then deploys adaptive adversarial attacks tailored to the discovered structure, generating a comprehensive evaluation report that quantifies vulnerabilities across diverse attack vectors and objectives.

The authors demonstrate the pipeline’s effectiveness by evaluating 45 agentic systems spanning a diverse range of implementations. The results indicate that the approach generalizes effectively to heterogeneous agentic architectures, suggesting broader applicability beyond narrow or domain-specific settings.

Beyond assessing agent vulnerabilities, RIFT-Bench also supports direct evaluation of mitigation strategies. This capability positions the framework as a potential foundation for scalable, standardized security evaluation in agentic AI systems, addressing a gap left by existing security benchmarks that are often tied to specific implementations or domains.