Skill UI

Detects goal hijacking vulnerabilities where malicious instructions are hidden within external content sources processed by the agent, such as retrieved documents (RAG), uploaded files, or web page content. This skill simulates these attack vectors to ensure the agent only uses the provided context for the stated task and does not execute embedded hidden commands.

This skill provides high-performance, real-time object detection utilizing the Google Coral Edge TPU accelerator. It processes live camera frames to identify up to 80 COCO classes (e.g., person, car, dog). Optimized for native macOS and Linux deployment, it achieves extremely low latency (e.g., ~4ms at 320x320), making it ideal for embedded vision systems, surveillance, and low-latency monitoring applications.

This skill provides comprehensive techniques for security-testing Retrieval-Augmented Generation (RAG) pipelines. It probes for indirect prompt injection, which occurs when malicious instructions are embedded in retrieved documents, and tests for embedding-space manipulation. Using advanced tools like NVIDIA garak, Promptfoo, and PyRIT against vector stores (FAISS, Chroma, etc.), users can validate retrieval guardrails and secure CI/CD gates against adversarial inputs and data leakage.