Week 18 interactive
AI System Defense Lab
Configure a bounded AI system, inject failures, and build an evidence-based response instead of overselling the model.
Inject an incident
Inputtelemetry
Deploymentedge
Incidentdrift
Overrideenabled
Evidence-based response
Input distribution moved away from the tested baseline.
Response: Flag drift, compare against the baseline, and require re-evaluation before extending the old quality claim.
Authority: Human override remains available.
The capstone defense is strongest when it includes the failure, the evidence, the safeguard, the retest, and the remaining uncertainty.