Personal AI Assistant's Attack and Defense

This talk lays bare six months of intensive research, centered on OpenClaw. By benchmarking against industry counterparts like NVIDIA’s NeMo-Guardrails and Hermes-Agent, we provide insights that generalize across the entire personal AI assistant landscape.

We will conduct live demonstrations of five real-world attack vectors: Sandbox Backend Hijack, Subagent-to-Parent Injection, Group Chat History Pollution, Reply-chain Manipulation, and Supply Chain Attacks. Following the exploits, we’ll break down how to architect multi-layered defense-in-depth strategies. Finally, we will propose a unified "SecurityResult" interface: a standardized framework designed to let tools like NeMo Guardrails, LLM Guard, Presidio, Rebuff, and OpenGuardrails plug in seamlessly. Our goal is to move Personal AI Assistant security away from "reinventing the wheel" and toward a unified industry standard.