c2.3 Establish AI security governance and a model inventory
Govern AI/ML and generative-AI systems before they scale: keep an inventory of models, agents, and data flows, assign ownership, and define acceptable-use and risk policies.
Implementation guidance
Keep a living inventory of AI in use
Maintain a single, current inventory of the models, agents, prompts, datasets, and third-party AI services in use — who owns each, what data it can access, and what risks it carries. Without it, shadow AI and ungoverned agents accumulate faster than security can review them.
Anchor decisions in established guidance
Frame your program around the AWS Generative AI Security Scoping Matrix, which classifies AI use cases by ownership of the model and data so you can choose controls that fit. Connect it to your existing risk, compliance, and data-classification processes.
Assign ownership and acceptable use
Name who owns AI risk decisions and define acceptable-use policies (what data may be sent to which service, which use cases are prohibited). Include embedded and third-party AI features, not just models you build, and revisit as agentic workloads grow.
Go deeper with the AI Security Maturity Model
This control is the foundational starting point for a Cloud program: establish the inventory, ownership, and acceptable-use basics. If AI is becoming central to your organization, the dedicated AI Security Maturity Model takes governance, data protection, and agentic-AI security through the full maturity journey — use it as the deep-dive companion to this control.
Related
Link this to Identify security and regulatory requirements and AI data protection.
How to validate
- Is there a current inventory of AI models, agents, and the data each can access?
- Who owns AI risk decisions, and how is acceptable use defined?
- Are third-party and embedded AI services included in the inventory?
Threats addressed
- Shadow AI and ungoverned agents create blind spots: data flows and model behavior no one reviewed, expanding the attack surface as AI workloads scale.