The Lethal Trifecta: A 2026 Defence Architecture for AI Agents
Read untrusted content, hold private data, act outward — pick two. The architecture rules that survive indirect prompt injection in production, with PromptArmor, classifier limits, and the trifecta principle.