mbhatt1/agentic_lpci_framework

LPCI is a security benchmark for testing language models against persistent, infrastructure-level prompt-control attacks. The main workflow configures a test run, has agents execute attack scenarios against selected model interfaces, then analyzes results and produces visualizations. A distinct infrastructure-attack path demonstrates poisoned retrieval, cross-session persistence, and tool-registry compromise. The excerpts directly show orchestration, RAG retrieval and session-store use, and analysis/visualization responsibilities; links for unsampled attack and service implementations are limited to documented capabilities and their listed paths.