Standard web access logs are not enough for AI systems. Prompt-level logging captures what an incident reconstruction actually needs.

A policy written before discovery fails. Find actual AI use through network review, audits, procurement records, and non-punitive surveys.

A generic IT incident response plan misses AI failure modes. A model-specific playbook covers rollback, evidence, and decision rights.

Raw model output should never reach a database, shell, or user unvalidated. Apply the same discipline used for user input.

Poisoning attacks corrupt training or fine-tuning data. Source control, provenance, and auditing reduce the risk before a training run.

Red teaming an AI system differs from red teaming a network. Define the adversary first, then test, document, and feed findings back.

An unverified model file is a supply chain risk. Provenance, signing, and version pinning bring model artifacts under control.

An AI agent should hold narrower permissions than the people who built it. Map its tools, scope its credentials, and review its grants.

Prompt injection cannot be fully solved at the model level, so layered controls must compensate.

AI security is a discipline with its own assets, adversaries, and failure modes. This guide maps the branch: what each control covers and where to start.