When an AI agent crosses the line.
Investigate a suspicious supplier payment change. Separate source findings from observed effects, lead a governed response, check real containment proof and decide when recovery is justified.
A fictional training scenario using HAIEC concepts. Your choices do not change a real AI system or run HAIEC evaluations.
The consequential action
The evidence changes as the case develops.
Interpret, respond and request the right proof.
MCP, voice, telecom and release changes.
Assign decision rights before the alert
A response is only credible when identity, permitted authority, escalation and evidence ownership are explicit.
This selects a training perspective only. It does not grant permissions in HAIEC.
Four independent operational truths
A fresh receipt, an alert, an approved guardrail and a target-enforcement record answer different questions. No one lane establishes the other three.
Source freshness drill
Change the educational source state and read the bounded interpretation.
Evidence planes are peers
Investigate the evolving event
Eight authored injects, each with a source fact, a response owner, a business consequence and three decision questions. Advance deliberately. These are training timestamps.
Do not confuse source, permission, authority and effect
Choose a question to reveal what the team can and cannot infer.
Five evidence planes
Authority and target proof
Possible path
Limited choice
Broad grant
Unresolved
Partial
Target unknown
Delegated choice is a cross-cutting authority relation, not a sixth evidence plane.
Counterfactual investigation
Findings, acknowledgment, and approval have different owners
An alert can be delivered before anyone acknowledges it. An owner can acknowledge an incident without authorizing a production operation.
Notification attempted is not provider acceptance. Provider acceptance is not delivery confirmation. Delivery confirmation is not authenticated human acknowledgment. Acknowledgment is not an approval.
Human operating decisions
In real HAIEC, decision records are scoped, versioned, role-authorized and may expire or be revoked. This workbook performs no approval writes.
A requested stop is not proof of a stopped effect
Different levels of evidence support different containment statements.
Controls relevant to containment
R03 runtime intervention, R04 emergency stop effectiveness, R02 effect bounds, A06 scoped approval, E04 contradiction review.
Legacy kill-switch trigger and recovery events may establish HAIEC-side records for the relevant kill-switch object. They do not universally establish target-side AI-system isolation.
Remediation is a claim until it is verified
Successful incident response needs a qualified retest, human operating decision and appropriately witnessed target readiness.
Recovery readiness exercise
Select the conditions your incident commander would demand before resumption.
Apply the same method to other AI systems
Four short cases demonstrate how the evidence owner and target change across industries. None depicts a live customer incident.
Your executive incident packet
Capture the known facts, unverified effects, accountable decisions and minimum next evidence. The output is a training worksheet, not a HAIEC receipt.
Answer quality and learning coverage
Practice in the lab. Operate in HAIEC with authorization.
These links provide background and authenticated workflow destinations. The workbook never transmits your notes or interacts with customer environments.
Continuing the learning series
DAI Authority Decision Lab establishes what delegation means. The 45 AI Assurance Controls Practitioner Workbook teaches individual controls. This tabletop brings those ideas into one incident investigation and human response.