Securing AI Systems
Short lessons for builders and security practitioners defending model-backed workflows. Each lesson teaches one idea in about a minute, using small fictional examples, so you can identify trust boundaries and enforce limited authority with defensive tests and evidence-based incident response.
Wulo Academy 74 lessons · 1 exercise · Self-paced · CertificateOverview
About this course
Short lessons for builders and security practitioners defending model-backed workflows. Each lesson teaches one idea in about a minute, using small fictional examples, so you can identify trust boundaries and enforce limited authority with defensive tests and evidence-based incident response.
Outcomes
What you will learn
Identify trust boundaries and enforce limited authority with defensive tests and evidence-based incident response.
Before you start
Requirements
Application and agent basics; all demonstrations require authorised isolated environments. Lessons build on each other; each lesson lists the earlier lessons it relies on.
Curriculum
74 lessons across 2 modules
01 First Section [edit me] 1 lesson
Test Beyond One Safe Response
02 Ungrouped 73 lessons · 1 exercise
Check Input and Output Separately
Cleaning Is Not Permission
Policy Must Control Action
Permissions Must Survive Traversal
A Work Directory Is Not a Sandbox
Keep Credentials Outside Context
Control Network Egress Separately
Keep Cache Reuse in Scope
Resist and Contain
Private Networks Need Clear Paths
Govern Every Tool Call
Map the Agent's Authority
Draw the Data Path
Likelihood Is Not Impact
Authorise Adversarial Testing
Different Jobs Make Better Layers
Reassess When Capabilities Change
When Capabilities Change
Name Who Accepts Remaining Risk
No Single Security Score
Instructions From Retrieved Data
When Systems Trust AI Output
Matching Words Is Not Meaning
Prompts Guide Gates Enforce
Trust the Source Not the Rank
Filter Before Building Context
Validate Files for RAG
Whose Authority Is the Tool Using?
Scope the Action and Target
Separate Inference from Authority
Two Ways an Agent Acts
Narrow Temporary Capabilities
Connection Is Not Permission
Narrow Tokens for Each Service
Budgets and Rate Limits
Upstream and Runtime Risks
Memory Crosses Session Boundaries
Keep User Memory Separate
Memory Does Not Grant Authority
Recover Beyond the Response
Map the Four Handoffs
When Routine Steps Combine
Do Actions Still Match the Goal
Permission Does Not Mean Appropriate
Identity Is Not Permission
Identity Is Not Permission
Prompts Are Not Tenant Isolation
Validate Upstream Models
Reassess Fine-Tuned Models
Signed Is Not Automatically Safe
Safer Model Artefact Loading
An SBOM Is Inventory Evidence
Stop Work and Revoke Access
When Permissions Outlive Work
Escalation needs a destination
Approve the Exact Action
Confidence Is Not Permission
Verify Through a Separate Route
Rehearse the Real Decision
Detection Is Not Proof
Recover from Classification Errors
Checks Before Delivery
Ownership Needs a Backup
Plan for Uncertain Futures
Make Findings Repeatable
Audit the Whole Data Path
Connect the Release Evidence
Boundaries for Security Logs
Preserve Evidence Before Cleanup
False Alarms Count Too
Policy Is Not Control Evidence
Policy Is Not Proof
Protect Training Coordination
Skills & tools
What you'll walk away with
- Securing AI Systems
Your instructor
Taught by a practitioner