Why it matters
NVIDIA's AI Kill Chain models attacks on AI applications as recon, poison, hijack, persist, impact, plus an iterate-and-pivot loop for autonomous agents. Each stage is paired with concrete controls and then applied to a RAG exfiltration path, connecting prompt injection to data ingestion, memory, tools, downstream actions, and monitoring.
My takeaway: Threat-model each route from attacker-controlled data to sensitive context or state-changing tools, then assign a preventive and detective control at every stage. Preserve data lineage, segregate trust domains, authorize each tool call against current user intent, gate persistent writes and scope changes, and instrument the chain so a partial bypass cannot silently become durable impact.