AI Data Poisoning: What Threat Hunters See

This report explains that AI data poisoning threats for most organizations originate upstream in vendors, third‑party packages, and CI/CD pipelines rather than directly in training datasets. It warns that agentic systems and LLMs used as evaluators create high‑value, high‑blast‑radius targets; emphasizes identity and dependency visibility; and recommends provenance verification, behavioral monitoring, and strict identity controls as first‑line detection and investigation signals while noting there is not yet a standardized incident playbook.