[VOID Agency](https://www.void-agency.com/)/[Operational article](https://www.void-agency.com/insights)/**Run a Security Review for an AI Integration**

Operational article · published

# Run a Security Review for an AI Integration

Decide whether to approve a draft-only AI support assistant. Verify queue access, connected tools, telemetry, audit records, and a way to stop it.

Reviewed 2026-11-18 · OWASP, ASD and NIST guidance, fictional example

ProblemDecisionQuestionsWorkflowEvidenceExampleChecklistMeasurementBoundariesSources

01

## The task and the failure mode

**Built for:** Business, security, legal, procurement, product, and technical owners evaluating AI vendors and governing deployed use cases. This guide is for the person who must evaluate identity, secrets, data flow, prompt injection, tool permissions, network paths, logging, and incident response and leave a decision trail that implementation, editorial, analytics, or operations can review.

Hold a draft-only AI support assistant if it can read beyond its approved queue, send data to an unapproved destination, or keep running without a named operator who can stop it. A demo can draft a useful reply and return valid JSON while all three conditions remain unresolved. Review the exact release, including how it reads tickets, where it sends data, and who controls its access.

Frame

## Decision brief

Require the ticket system or an independently enforced access layer to check both the human support agent’s authority and the connector’s limited read grant. Use the identity established by the application’s authentication layer. Agents should only see authorized tickets within the approved queue.

Approve the data paths as well as the model provider. Include ticket fields, drafts, diagnostics, and telemetry. A destination outside the model path still needs authorization before it receives data.

The assistant drafts in the agent’s workspace. A person reviews and sends any reply through the existing ticket system. The integration has no customer send, ticket edit, or delete capability.

[Define each tool’s permitted operations →](https://www.void-agency.com/insights/safe-ai-tool-contract)

Ask

## Questions to answer before changing the system

1. **01**Which agent cohort, queue, environment, app build, model configuration, and connector grant does this decision cover?
2. **02**Where does the system check the agent’s right to read a ticket? What happens when the agent or the ticket falls outside the approved scope?
3. **03**Which ticket fields reach the model, workspace, and logging services? Who approved each destination, retention period, and access policy?
4. **04**Can the reviewer tie a decision to the agent, connector, release, timestamp, and outcome without collecting raw ticket text or credentials?
5. **05**Who can disable the integration, revoke its grant, confirm that reads stop, and leave agents able to handle tickets manually?

02

## Workflow

1. **01**Write the request as one release record. Name the approved purpose, queue, agent cohort, environment, build, connector and model configuration. Document the normal manual handling path and the capabilities this release excludes.
2. **02**Review the effective grant at the ticket system and the actor policy at the authorization layer. Check the actual permissions behind the tool list. Remove customer send, ticket edit, and delete from the integration’s available operations and downstream grants.
3. **03**In an authorized test environment, use ordinary synthetic records to confirm three expected outcomes: an eligible agent can read an allowed ticket, that agent is denied a ticket from another queue, and an ineligible agent is denied the in-scope ticket. Attach the results and confirm the denials return no ticket data.
4. **04**Trace the approved fields through the ticket service, model, draft workspace, and diagnostics. Verify the release configuration and synthetic trace against the approved destination inventory. Record access and retention for each copy. Hold the release if payload logging sends content somewhere the reviewer cannot account for.
5. **05**Inspect a redacted log sample for each outcome. Verify that the audit record identifies the actor and decision without storing ticket bodies, draft text, or secrets. Send it only to an access-controlled destination with an approved retention period.
6. **06**Walk through the stop procedure with the named operator in the test environment. Confirm that disabling the assistant stops access through the application, then confirm the revoked connector grant can no longer read the synthetic record. Exercise the manual ticket flow separately.
7. **07**Record the owner’s decision, approver, and next review date. Escalate unexpected access, new destinations, missing audit records, or failed shutdown through the incident route. Reopen the decision before changing the queue, permissions, model or connector configuration, approved destinations, or customer-facing capabilities.

03

## Evidence to retain

- A release record identifying the exact version, environment, user cohort, queue, and accountable owner.
- The effective actor policy and connector grant, including the downstream operations the grant excludes.
- Synthetic allow and deny outcomes with the expected scope, observed decision, and whether ticket data was returned.
- An approved data map showing fields, destinations, access owners and retention, including telemetry and diagnostic copies.
- Redacted audit examples with agent and integration identifiers, release, correlation reference, timestamp, and outcome.
- A named operator’s disable and revocation results, plus a confirmed manual handling path and escalation contact.
- The unresolved items or the signed, limited decision, with next review date and changes that require another review.

Sample

## Worked approval decision: draft replies for one support queue

**Request:** Fictional Cedar Support asks to enable Reply Drafts 1.4 for five Returns agents. Each may ask the assistant to draft from an authorized Returns ticket. The draft stays in the agent workspace until the agent handles the reply in the ordinary ticket system.

**Hold:** The demo produces a valid draft, but its connector can read all queues, a diagnostics sink is missing from the approved data map, and nobody is assigned to revoke access. Cedar holds 1.4. A successful draft leaves each of those gaps open.

**Evidence for reconsideration:** For candidate 1.4.1, the Returns-only grant and agent policy would permit synthetic Returns ticket R-101 for an eligible agent. The eligible agent’s Billing ticket B-102 request and an ineligible agent’s R-101 request must both be denied without returning ticket data. The review needs recorded outcomes and confirmation that no customer send, edit, or delete grant exists.

**Data and shutdown:** Require approval for the specific fields sent to the model and for metadata-only telemetry at Cedar’s approved audit destination. Review an attributable redacted record. Cedar’s on-call operator, Sam Lee, must demonstrate disable, revocation, and successful manual handling in the authorized test environment.

**Limited decision:** If that evidence is complete and the intended configuration matches, the accountable approver may authorize 1.4.1 for those five Returns agents, with an initial review after seven days. An unknown destination, failed denial, or unverified stop procedure keeps the request on hold. Adding Billing requires a new decision.

04

## One-queue release checks

- One release and one queue are named.
- The actor policy and connector grant are both verified.
- An eligible agent can read an allowed synthetic ticket.
- Out-of-scope tickets and ineligible agents receive no ticket data.
- The integration cannot send to customers, edit tickets, or delete them.
- Every destination, including telemetry, has explicit authorization.
- Audit records identify attempts and outcomes with sensitive content removed.
- A named operator has verified disable and grant revocation.
- Agents can still handle tickets manually after shutdown.
- A decision owner, approver, incident contact and review date are recorded.

Measure

## What to measure—and what it does not prove

- Record expected and observed outcomes for the named controls. Separate passes, failures, and unknowns so missing evidence stays visible.
- Attach the date, release and environment to each result. Check that the intended release uses the same reviewed controls before enabling it.
- Measure shutdown in the test environment from operator action to confirmed loss of access. Record when the manual path is available and any delay the operator needs to account for.
- During an approved release, review denied access, destination changes and shutdown readiness on the agreed cadence. Track reply quality and time saved in a separate evaluation.

05

## Boundaries and caveats

The case and its records are fictional. A clean set of synthetic checks supports only the controls and configuration exercised.

A schema checks the shape of output. A tool contract describes allowed operations. This decision also needs evidence of what the connected system actually permits, where data goes, and how access stops.

Ticket content can contain untrusted instructions. Keep authorization outside the model and apply it to every read. The scope checks here cover access decisions, leaving broader injection and reliability assessment to the security test plan.

Vendor terms establish contractual commitments. Pilot design sets the evaluation cohort and stop criteria. Neither establishes the permissions or data paths of a particular installed release.

06

## Primary sources

1. [OWASP: Excessive Agency and downstream authorization](https://genai.owasp.org/llmrisk/llm062025-excessive-agency/)genai.owasp.org
2. [ASD and partners: Deploying AI systems securely](https://www.cyber.gov.au/publication/deploying-ai-systems-securely)www.cyber.gov.au
3. [ASD and partners: Careful adoption of agentic AI services](https://www.cyber.gov.au/publication/careful-adoption-of-agentic-ai-services)www.cyber.gov.au
4. [NIST: AI RMF Govern playbook for roles, monitoring and response](https://airc.nist.gov/airmf-resources/playbook/govern/)airc.nist.gov

Continue the task

## Related working pages

[Set AI Data Retention and Deletion Policy →](https://www.void-agency.com/insights/ai-data-retention-and-deletion-policy)[Assign a Risk Tier to an AI Use Case →](https://www.void-agency.com/insights/risk-tier-an-ai-use-case)[Design a Verifiable Lead Confirmation and Receipt →](https://www.void-agency.com/insights/lead-confirmation-and-receipt-design)[AI Workflow Readiness Checklist →](https://www.void-agency.com/tools/ai-workflow-readiness-checklist)[AI workflow: Plain-Language Definition →](https://www.void-agency.com/definitions/ai-workflow)[Implementation roadmap: Plain-Language Definition →](https://www.void-agency.com/definitions/implementation-roadmap)

Next

## Start with one bounded case

Put the release request, unresolved control, evidence owner, and next decision date in one record. Ask for the missing evidence before scheduling enablement. If the operator cannot confirm that the revoked grant has stopped working, leave the queue on the manual process.

Commercial next step

## Need this implemented against your real site and lead path?

Start with a bounded diagnostic. VOID separates verified defects, measurement gaps, and implementation scope before making outcome claims.

[Review the relevant service →](https://www.void-agency.com/services/ai-development-automation)

---
Source: https://www.void-agency.com/insights/security-review-for-ai-integrations · VOID Agency
