This repository provides `agentic_security`, an open-source vulnerability scanner for Large Language Models (LLMs) and agent workflows. It is a FastAPI web application with a Vue.js frontend, a Python CLI (`agentic_security`), and a separate policy-driven agent harness CLI (`agesec`). Its primary function is to red-team LLM endpoints by sending adversarial prompts (jailbreaks, prompt injections, multimodal attacks) and classifying the responses for security failures (refusals, data leaks, compliance). The architecture is a monolith with well-defined modules: a core scanning engine (`probe_actor`, `fuzz_chain`), a dataset loading and transformation system (`probe_data`), a pluggable refusal classifier (`refusal_classifier`), and an LLM provider abstraction layer (`llm_providers`). It also contains a set of highly reusable async resilience components (`executor/`) and a durable event-sourced run log (`agentic_security_harness/runlog.py`). While primarily an application, it has strong data/research characteristics. It ingests attack datasets from Hugging Face, local files, and dynamic generators. It includes trained sklearn models for refusal classification, RL-based prompt selection, and steganographic prompt transformations. This hybrid nature makes it a rich tool for both security engineers and AI safety researchers. The codebase is actively maintained but exhibits significant technical debt. This includes an unfinished architectural migration (removing MCP/Agno), syntax errors in core files, hardcoded secrets, an aggressive Python version requirement (>=3.15), and bundled outdated frontend dependencies (Vue.js 2.6.12). Despite these issues, the core scanning logic and several reusable components demonstrate high craftsmanship, making it a valuable reference for building LLM security tooling.