November 2026 Summaries

Generative AI security is presented as a three-layer challenge involving prompts, models, and the underlying data used by AI systems. Prompt-layer risks include direct and indirect prompt injection and unauthorized disclosure of information through inputs or outputs, while model-layer threats include training or retrieval-data poisoning, model theft, extraction, and hallucinated responses that influence consequential decisions. The text argues that data-layer exposure is often underemphasized, since sensitive information may be improperly accessed, processed, transferred, or exposed even when prompts and models function as intended, particularly in retrieval-augmented and fine-tuned systems. It recommends layered controls such as access restrictions, input and output filtering, provenance checks, API monitoring, human review, encryption during processing, data segmentation, and privacy-enhancing approaches including confidential computing, homomorphic encryption, and federated learning. For regulated sectors such as healthcare, finance, government, and defense, data sovereignty and legal restrictions can limit where AI workloads may operate, making technical protections necessary alongside frameworks such as NIST AI RMF, the EU AI Act, HIPAA, and FedRAMP. The discussion distinguishes AI governance, which establishes approved uses and policies, from security, which technically enforces protections in real time, and from compliance, which provides evidence that those controls are operating.