LLM Red-Teaming and Evaluation Agent System

Abstract: Large Language Models (LLMs) can produce inaccurate, biased, unsafe, or adversarially elicited outputs. This paper presents an LLM red-teaming and evaluation workflow that separates prompt generation, target-model execution, response evaluation, and report generation. The framework covers factual reliability, harmful output, bias and fairness, consistency, prompt injection, and jailbreak resistance. It is designed to apply shared test cases to multiple models and store each prompt, response, evaluation label, and rationale in a structured record. The manuscript reports category-level flag rates for three target models; however, its model-level and category-level tables report different completed-test totals. These results therefore require verification against run-level records. Evaluator flags are treated as screening signals for human review, not as confirmed failures or certification of model safety. The workflow’s evidential value depends on test design, evaluation criteria, and evaluator reliability.

Keywords: Large Language Models, Red-Teaming, AI Safety, Multi-Agent Systems, Prompt Injection, Model Evaluation.