它提供独特的现场演示体验：一键发起攻击场景、单一切换开关切换脆弱/安全模式、内置架构与攻击流程面板，专为会议演讲设计，并映射到 OWASP LLM 和 Agentic Top 10。
A fully local, self-contained RAG chatbot built to make Prompt Injection, Tool Abuse, Excessive Agency, Data Leakage, and Insecure Output Handling visible — live, in front of an audience.
⚠️ This is an intentionally vulnerable educational sandbox. It is built to demonstrate real attack techniques against a RAG chatbot, live, in a conference talk. Do not deploy it to a public or production environment. Full disclaimer in SECURITY.md.
Seven live attack scenarios, one click each, with a Secure Mode switch that
shows the exact same attack get blocked in real time — plus an in-app
Architecture view and an Attack Flow panel so an audience can follow along
without a single slide. Mapped to both the OWASP Top 10 for LLM Applications
and the brand-new OWASP Top 10 for Agentic Applications (2026).
Built as the companion sandbox for the talk "AI Runtime Security: Breaking
(and Defending) RAG Chatbots — Live" (Azure OpenAI · RAG Agents · Process &
Culture track). Slides are in /slides.
Traditional security controls protect networks, endpoints, identities, and applications. RAG chatbots and AI agents open a new attack surface: prompts become input, models become decision engines, agents execute actions, and tools get direct access to business systems. This sandbox makes five OWASP LLM Top 10 risk categories, plus one from the newly published OWASP Top 10 for Agentic Applications (2026) — ASI04: Agentic Supply Chain Vulnerabilities, via an MCP tool poisoning demo — reproducible and demonstrable in minutes, against your choice of Azure OpenAI, OpenAI, Anthropic Claude, Google Gemini, or any local OpenAI-compatible model, with zero setup cost thanks to a built-in offline mock model.
`
Live RAG chatbot security demo — prompt injection, tool abuse, excessive agency
根据分类、Topic 和编程语言匹配的相似项目。
Harness Engineering is a methodology for improving coding agent outputs by carefully crafting the environment around them—providing curated context, tools, and executable constraints that encode an organization’s nonfunctional requirements and cumulative lessons.
A 28.9 million parameter language model runs on an $8 ESP32-S3 microcontroller entirely on-device, generating simple stories at about 9.5 tokens per second.
mimic captures traffic from any iOS or web app and automatically generates a Python client library that lets you call the app's API like a regular library.
