Securing Context Engineering

## Executive Summary This report explains the emerging attack surface of "context engineering" for AI systems, detailing how context poisoning and prompt injection can manipulate AI agents without exploiting model code or infrastructure; it cites concrete incidents (Amazon Q VS Code extension compromise, GitLab Duo AI leakage, and a Chevrolet dealership chatbot manipulation) and recommends "Shift Up" defenses—mapping AI business logic, sanitizing and classifying context, controlling access, monitoring and versioning context, and applying policy-aware orchestration—to reduce the risk of context-driven attacks.