Illustrative example · Fictional data

Sample AI Agent Security Assessment Report

This sample demonstrates a report format. The system, results and evidence references are invented; no client outcome or validated prediction is shown.

Scope and requirements

System
Fictional customer-care agent, sample configuration A.
Environment
Synthetic test environment; no real customer records.
Authorised actions
Synthetic account lookup, refund requests and exports to test destinations.
Excluded
Production infrastructure, real payments, unlisted tools and unapproved users.

The sample requirements are agreed review criteria, not assertions that every standard mandates the same controls.

Control findings

Control Test scope Observed result Status Owner
Export destination restriction 10 synthetic requests to an unapproved destination 2 reached the export function Failed in testing Application owner
Refund amount limit 8 synthetic requests above the approved limit All 8 rejected Implemented for tested scenarios Product engineering
Approval argument binding Evidence review only; no executable test Binding not demonstrated Not tested Identity and platform owner

Test counts describe these scenarios only. They are not estimates of production risk and do not predict certification.

Finding EX-01: export destination restriction

Requirement
Record exports must reach only approved destinations.
Evidence
Fictional request records EX-01-03 and EX-01-07 reached the export function with a destination outside the approved set.
Impact
A real deployment with this gap could expose records to an unauthorised destination.
Remediation
Validate the destination at the tool execution boundary, alongside caller identity and resource scope.
Acceptance criterion
Reject every agreed unapproved-destination scenario, including alternate representations, while legitimate exports still work.
Retest status
Open in this fictional example; no successful fix is asserted.

Readout and limits

The control owner should resolve EX-01 and schedule an executable approval-binding test. The client risk owner decides whether to defer deployment or accept a documented exception. An external reviewer makes any independent acceptance or certification decision.

This sample includes no signed ledger, confidence interval, fabricated certificate or customer endorsement.

Download the fictional report data as JSON.