Playbook: AI Security & Threat Modeling

This playbook teaches teams how to find attacker surfaces in AI workflows, choose practical mitigations, and build operating controls so models and integrations run with clearer security boundaries.

Why this matters

AI components change where and how data, models, and decisions flow. A well-scoped threat model helps you identify risks before code ships: data leakage from prompts or logs, unauthorized model queries, supply‑chain exposure from third‑party models, and runtime failures that reveal sensitive inputs. Left unchecked, these problems create operational, legal, and reputational harm across small teams, service businesses, healthcare providers, manufacturers, research groups, and public institutions.

What you will understand and be able to do

Using this playbook you will learn to:

• Map AI assets and data flows (inputs, models, outputs, stores, and integrations).
• Create simple threat models for common AI use cases (customer chatbots, document automation, predictive maintenance, and clinical decision support).
• Choose and apply authentication, authorization, and least‑privilege patterns for model access and APIs.
• Design data handling controls: minimization, masking, encryption, and logging policies that balance utility and privacy.
• Define monitoring, observability, and alerting for model performance, anomalous queries, and cost spikes.
• Develop practical mitigations and runbooks for incidents, plus a cadence for tabletop tests and audits.

Who benefits

This playbook is intended for engineering, IT, security, SRE, and product teams supporting AI features — and for managers, consultants, and operators who must make deployment decisions. Examples: a neighborhood clinic protecting patient inputs to a triage model; a small e‑commerce team securing a recommendation API; a manufacturer locking down predictive maintenance telemetry; and a university lab preparing a safe model release.

How this fits the Applying AI domain

Applying AI focuses on practical paths from idea to impact. This playbook supplies the security and operational discipline that keeps those paths open. Use it alongside resources on automation, decision support, and organizational readiness to move from “can we?” to “how do we safely and sustainably?”

Practical next steps

Start by inventorying your AI use cases and the data each one touches. Run a short tabletop to expose assumptions, then build a lightweight threat model and prioritized mitigation list. Use accessible controls first: API keys scoped to roles, input/output logging with retention rules, encryption in transit and at rest, and basic anomaly alerts. Iterate: test mitigations, run adversarial checks, and update runbooks as you learn.

Platform opportunities: consider saving assessments with the playbook's interactive forms, or copy and tailor this playbook into your team space so policies, checklists, and audits evolve with your systems.