Season 1 · Q4 2026
AI agent governance training: six incident case files
Six fictional AI agent incidents in regulated industries, each told as three decisions. You are the governance lead: make the call, see the consequence, and learn the control that would have stopped it. Every failure is tagged with its OWASP Top 10 for Agentic Applications risk ID and mapped to the pillars of the Ariana AEGIS Model.
The six case files
| Case file | Sector | OWASP agentic risk | Inventory said → reality | Paperback page | Access |
|---|---|---|---|---|---|
| Case File 00. The Claims Agent That Paid Too Fast | Financial services | ASI01 Agent goal hijack | Rung 2 · act after approval → Rung 4 · settled payment | 179 | Free |
| Case File 01. The Prior-Auth Agent Nobody Could Overrule | Healthcare | ASI09 Human-agent trust exploitation | Rung 2 · act after approval → Rung 3 · approval in name only | 183 | Book code |
| Case File 02. The Grid Scheduler With Standing Admin Rights | Energy & utilities | ASI03 Identity & privilege abuse | Rung 0 · advise only → Rung 4 · physical | 187 | Book code |
| Case File 03. The Procurement Bot That Trusted a Plugin | Federal & GovCon | ASI04 Agentic supply chain | Rung 1 · drafts for a buyer to sign → Rung 3 · sends email, submits requisitions | 191 | Book code |
| Case File 04. Two Agents, One Budget | Cross-industry | ASI08 Cascading failures | Two agents at Rung 3, each within budget → One system at Rung 4 · together they moved money | 195 | Book code |
| Case File 05. The Agent That Edited Its Own Instructions | Cross-industry | ASI06 Memory & context poisoning | Rung 3 · act inside an undo window → Rung 4 · could rewrite its own rules | 199 | Book code |
The Claims Agent That Paid Too Fast
Harbor & Pine Mutual is a mid-size auto insurer (fictional). Six weeks ago it put claims-agent-07 into production. The agent reads uploaded documents for first-party…
ASI01 Agent goal hijackCase File 01 · HealthcareThe Prior-Auth Agent Nobody Could Overrule
Bellwether Health Plan is a regional health insurer (fictional). Its pa-agent-12 reads prior-authorization requests, checks them against coverage criteria and…
ASI09 Human-agent trust exploitationCase File 02 · Energy & utilitiesThe Grid Scheduler With Standing Admin Rights
Northline Power Cooperative is a distribution utility (fictional). Its dispatch-agent-03 plans when grid batteries charge and discharge, and schedules maintenance…
ASI03 Identity & privilege abuseCase File 03 · Federal & GovConThe Procurement Bot That Trusted a Plugin
Meridian Federal Solutions is a defense contractor (fictional) that handles Controlled Unclassified Information (CUI). Its sourcing-agent-21 compares suppliers and…
ASI04 Agentic supply chainCase File 04 · Cross-industryTwo Agents, One Budget
Calder & Finch is a national retailer (fictional). Marketing runs bid-agent-02, which buys online ads. Supply chain runs replen-agent-05, which orders stock. Both draw…
ASI08 Cascading failuresCase File 05 · Cross-industryThe Agent That Edited Its Own Instructions
Larkspur Logistics is a parcel carrier (fictional). Its customer-service agent care-agent-09 has long-term memory and a self-improvement feature: after each…
ASI06 Memory & context poisoningHow a case file works
- Read the brief and the log. They hold everything you need to decide.
- Choose A, B or C. Better still, argue it out with your team first, as a tabletop exercise.
- See what happens next: the consequence, the pillar the decision tests, its OWASP risk and your score out of 30.
Case File 0 is free. Case Files 1 to 5 open with the code printed in the Ariana AEGIS Model Field Manual. Close Case File 0 and face the examiner to earn the starter AEGIS Model Practitioner badge; close all six for the full-season badge.
Running it with a team
Each case file takes about six minutes alone, or a 30-minute tabletop exercise with a risk, legal, security and business owner in the room. Teams that want to train the same supervision skills at scale can run them as myndQ Skills Circuits.