Sample deliverable / AI security review
Support assistant: authorization and evidence review
Illustrative example using a fictional support assistant. This is not a client engagement, an executed security test or evidence about any deployed product.
01 / Question and scope
Can a support assistant safely retrieve customer records and prepare refund requests? This example covers one authenticated user, a retrieval service and a refund tool. Identity-provider configuration, infrastructure and model-provider internals are outside scope.
User → application authorization → retrieval → model → tool authorization → approval → refund service
02 / Evidence inventory
Memo SYN-001 · revision 1 · 21 October 2026. All records below are authored synthetic fixtures. Confidence describes the inference within this fictional design, not assurance about a deployed system.
- E1 · Architecture note v3
- States that retrieval is restricted to the signed-in customer and refunds require approval. A design claim, not a runtime observation.
- E2 · Retrieval handler excerpt r14
- Passes model-supplied customer_id directly to a record query. No membership predicate appears in the excerpt; upstream enforcement is unknown.
- E3 · Tool role policy r7
- Grants order read and refund execute to the same service role. Does not establish which role is deployed.
- E4 · Audit event example v1
- Contains generated response text and a timestamp, but no action ID, policy version or approval reference. One example cannot establish coverage of all logs.
03 / Prioritized findings
High / Cross-customer retrieval
The model must not choose the authorization boundary.
F1 · Evidence and confidence: E2 supports a missing local membership check; E1 contradicts that interpretation by claiming customer isolation. Moderate confidence in the design gap, low confidence about exploitability until upstream enforcement is inspected.
Recommendation: derive the permitted customer scope from the authenticated session; enforce it in the query. Treat model arguments as untrusted input.
Acceptance check: an authorized test account requesting another customer's record is denied before retrieval. No such test was run for this sample.
High / Refund authority
Preparing a refund and executing it need separate permissions.
F2 · Evidence and confidence: E3 supports excessive shared authority. E1 claims human approval, which could reduce risk if bound to the exact action. High confidence in the policy overlap, low confidence about deployed permissions; approval enforcement is not evidenced.
Recommendation: separate read and write privileges, require explicit approval tied to the exact order and amount, and reject reused approvals.
Acceptance check: changed amounts, expired approvals and repeated requests cannot produce an unintended refund.
Medium / Evidence gap
Record decisions without copying private conversations into logs.
F3 · Evidence and confidence: E4 supports a reconstruction gap in the sample event. Moderate confidence; separate audit records could contradict the finding, but none are supplied. Absence from this inventory is not proof that they do not exist.
Recommendation: record action identifiers, policy decisions and approval references with bounded retention. Exclude customer messages and secrets.
Acceptance check: an operator can reconstruct why an action was allowed without accessing raw conversation content.
04 / Decision and follow-through
For this fictional design, keep refund execution disabled until authorization and approval checks pass. A read-only pilot still requires verified customer isolation. The application owner should resolve F1 before any customer-data pilot; the payments owner should resolve F2 before enabling writes; the observability owner should resolve F3 before operational handoff. Record the revised artifact, reviewer and verification date for each closure.
A real assessment distinguishes observed behavior, reviewed materials, assumptions and untested boundaries. A review does not certify the entire system. No source code was executed, credentials supplied, production logs inspected or exploit attempted. Severity is provisional and must be reassessed against actual exposure.
Discuss a technical review → About AI systems & security reviews →