Domain 4 of 4 · Chapter 2 of 3

Secure AI and SAIF

Unlock the complete study guide + 1,040 practice questions across 16 full exams.

Bundled into the existing Generative AI Leader premium course — no separate purchase.

14-day money-back guarantee — no questions asked.

Included in this chapter:

  • What secure AI means, and why the whole lifecycle
  • The AI threat surface across the lifecycle
  • SAIF: purpose, the Map, and the Risk Self-Assessment
  • Google Cloud security tools, and the need each serves

Which Google Cloud security tool for which need

Decision axisIAMSecurity Command CenterSecure-by-design infrastructureWorkload monitoringModel Armor
Primary jobControl who can access whatSee and triage security postureBuild protection into the platformRecord and flag activityScreen live prompts and responses
Lifecycle surface it guardsData and Model accessWhole environmentInfrastructure foundationRunning workloadsApplication (model input/output)
Example AI risk addressedModel exfiltration, unauthorized data accessExposed or misconfigured AI assetsTampering, unauthorized low-level readsUnnoticed anomalous or malicious activityPrompt injection, jailbreak, data leakage
What it producesRoles and permissions on resourcesFindings, alerts, AI asset inventoryEncryption, isolation, secure defaultsAudit logs and detectionsA sanitized or blocked prompt or response
Mostly built in or configuredCustomer configures grantsCustomer enables and reviewsGoogle builds in; customer tunesOn by default; customer reviewsCustomer enables per application

Decision tree

Fairness or explainability?Responsible AInot a security toolRestrict who can accessmodels or data?IAMroles, least privilegeScreen a live LLM'sprompts or responses?Model ArmorCentral view ofexposed AI assets?Security Command CenterCloud Audit Logsworkload monitoringYesNoYesNoYesNoYesNoAlways on beneath all of these: secure-by-design infrastructure
Cheat sheet

Cheat sheet

  • Secure AI protects the whole ML lifecycle, not just the model
  • SAIF is Google's framework for building AI secure by default
  • SAIF splits an AI system into Data, Infrastructure, Model, Application
  • The SAIF Map pairs every AI risk with a control that mitigates it
  • Use the SAIF Risk Self-Assessment to scope a project's risks before launch
  • Data poisoning corrupts training data so the model learns the wrong thing
  • Model exfiltration is theft of the trained model itself
  • Prompt injection hides malicious instructions in a live model's input
  • Rogue actions are unintended or malicious steps an AI agent takes
  • IAM controls who can do what on which resource
  • Security Command Center is the central posture and threat-detection service
  • Secure-by-design infrastructure builds protection into the platform itself
  • Workload monitoring uses Cloud Audit Logs to record who did what, when
  • Model Armor screens a live LLM's prompts and responses
  • Secure AI layers the tools; use them together, not one instead of another
  • Place an AI security threat at the SAIF lifecycle stage where it strikes

Unlock with Premium — includes all practice exams and the complete study guide.

References

  1. Google Secure AI Framework (SAIF)
  2. SAIF components
  3. SAIF risks (risk map)
  4. SAIF Map
  5. SAIF home
  6. Secure AI Framework (SAIF) on Google Cloud
  7. Google Cloud security for AI
  8. IAM overview
  9. Security Command Center
  10. Cloud Audit Logs
  11. Model Armor overview