[NAACL2024] Attacks, Defenses and Evaluations for LLM Conversation Safety: A Survey

Agentic chatbot user simulation workflow to generate diverse and realistic personalities with conversations. Tests chatbots across multiple dimensions (off-topic handling, hallucination detection, safety, prompt extraction, financial advice) using a locally hosted open source LLM. Features a simple webapp, CLI tools, SQLite database storage.

LLMrv is a framework for monitoring LLM conversations against formal safety policies in real time. It models conversations as event traces, specifies policies in past-time temporal logic (ptLTL), and bridges the gap between formal Boolean semantics and free-form natural language through a semantic grounding layer.

Making LLM conversations public for AI Safety research

CrisisBench evaluates LLMs for their safety in crisis conversations

SafetyDriftBench: measuring LLM safety-rule drift over long conversations

LLM-based mental health chatbot with RAG, LangChain, and safety filters for empathetic conversations.

AI-powered Health Assistant with Streamlit & Local LLM, multi-turn conversation, and safety-first responses

Standalone benchmark for multi-turn safety persistence in medical LLM conversations. Measures recommendation monotonicity under sustained patient pressure.

Notes from 100+ hours of stress-testing LLM conversations, focusing on tone, trust, emotional safety, and linguistic nuance

Topic modeling pipeline for LLM safety monitoring using BERTopic and local LLMs (Ollama). Includes two-pass labeling with chain-of-thought reasoning, confidence scoring, and comprehensive visualizations for identifying safety-critical conversation clusters.

AI Health Chatbot using LLM (Llama 3.1 8B) | Multi-turn Conversation | Safety Filters for Emergency & Crisis | HuggingFace Router API

A scalable system for evaluating conversation turns on hundreds to thousands of linguistic, pragmatic, safety, and emotional facets using open-weight LLMs.

AI Agent system for analyzing doctor–patient conversations using LangGraph. Automatically routes transcripts, extracts structured drug prescriptions with LLM-powered tools, validates safety, and returns auditable clinical insights.

A high-performance chat engine built with FastAPI and Google's Gemini LLM, designed to handle concurrent conversations with sophisticated thread safety and state management. The system implements advanced async patterns to process multiple chat streams while maintaining conversation context and handling complex mathematical operations.

SentinelLM is a proactive AI defense system that protects LLM applications from prompt injection attacks and misuse. By analyzing prompts in real time, classifying risks, and enforcing safety policies, it acts as a firewall for AI — keeping conversations secure, reliable, and abuse-free.

Text and voice-based AI Assistant built with LiveKit Agents SDK and OpenAI GPT-4o. Supports real-time conversations, tool calls (weather, promptify, embeddings), grounding & safety checks, and CI-tested with pytest. Provides both CLI chat and voice-ready interfaces for scalable LLM agent development.