StackPilot / Occupation guides / Manual 026

Occupation manual 026 · review-only workflow

AI can organize design-assurance labels. Security-engineering decisions stay human.

Use AI to preserve approved reference labels, owners, times, stated purposes, and unknowns. Do not let it select requirements, assess threats, choose controls, generate test results, accept risk, change a system, deploy, or take external action.

For: security-engineering teamsOutput: design-assurance packetReviewed: 2026-08-01
Answer-first field map

Build a review packet, not a security-engineering result.

Information Security Engineers develop and oversee security procedures and policies, build or upgrade security technology, design and implement controls, and may oversee assessment.1 Systems security engineering spans security requirements, design, verification, validation, and assurance across a lifecycle.2 Those are consequential activities. This safe workflow only organizes fictional or organization-approved minimum-necessary labels and hands unanswered questions to named authorities.

InputApproved labels only: supplied reference, owner, time/version, stated purpose, type, and known gaps.
OutputA Design-Assurance Packet ending in needs_review.
Not an outputNo threat, vulnerability, requirement, control, test, finding, risk, exception, change, deployment, or message.
Standards-source limit: NIST systems-security, secure-development, and AI-risk materials are context. They are not your engineering method, authorization path, approved tool list, test plan, control catalog, risk rubric, or permission to expose security material to a model.
  1. Set a strict scope before any AI work.

    Give the packet a fictional or approved scenario label, a stated purpose, and a named authority owner. Exclude system names, code, diagrams, configurations, credentials, identities, ticket text, test output, findings, security-tool output, customer data, and other private/security-sensitive material.

  2. Register supplied labels without turning them into evidence.

    For every supplied label, record its reference, owner, time/version, stated purpose, and type: requirement, control, test, or decision. Use only supplied, unknown, or conflicting status. Never ask a model whether a design is secure or what it should do.

  3. Let AI organize, not inspect or engineer.

    AI may format supplied labels, flag blank fields, group duplicates, and draft neutral questions. It must not inspect a system or repository, infer a threat/vulnerability, select a control, create a test result, or draw a security conclusion.

  4. Preserve gaps as reviewer questions.

    For every missing owner, time, source, approval, policy, or lifecycle gate, retain unknown. Ask the named security-engineering owner which current process governs the missing fact; do not fill it from model inference.

  5. Route authority to the right owner.

    Security engineering owns technical assessment; technical and system owners own implementation context; privacy, legal, and compliance own obligations; procurement owns vendor approval; risk owners accept risk; operations owns execution; records owners govern retention.

  6. Stop at the review handoff.

    Create only a local needs_review receipt. Do not issue a finding, edit a ticket, change code/configuration, select a requirement/control/test, approve an exception, deploy, purchase, communicate, or take any external action.

Narrow agent roles

Three bounded roles. One human control point.

Reference stewardNormalizes supplied reference labels and flags blank owner, time, purpose, type, or scope fields. No retrieval, interpretation, or assessment.
Unknown clerkGroups duplicate or contradictory labels and marks them unknown or conflicting. No requirement, control, test, or risk conclusion.
Handoff assemblerFormats neutral reviewer questions and a local receipt from approved labels. No system connection, ticket, approval, change, deployment, or external output.

NIST's SSDF describes high-level secure-development practices that can be integrated into each SDLC implementation.3 A label packet is not a completed SSDF practice, a passing test, or a conforming security program.

Safe versus approval-required

Let AI prepare labels. Let authorized people control engineering.

AI may prepareNamed human approval is required
Format supplied fictional or approved minimum-necessary labels.Access, inspect, retrieve, upload, query, or analyze any system, repository, security tool, test environment, ticket, design, configuration, or record.
Preserve supplied reference, owner, time/version, stated purpose, type, and unknown/conflict labels.Identify a threat or vulnerability; select/approve requirements, architecture, controls, configurations, tests, findings, exceptions, risk, residual risk, or readiness.
Draft neutral questions about missing authority or contradictory labels.Claim verification, validation, compliance, authorization, or assurance; change code, a control, configuration, credential, record, or ticket; deploy; procure; or communicate.
Produce a local review-only receipt.Approve a vendor/tool, make legal/privacy/compliance conclusions, accept risk, or take any external action.
Stop immediately if the request includes private system/design/test material or asks for a security judgment, control choice, test conclusion, risk decision, exception, change, deployment, communication, or external action.
Failure modes

Five ways label prep quietly becomes unauthorized engineering.

MistakeWhy it failsRepair
“Choose the right control.”Control selection is an engineering and risk decision.Preserve the supplied labels and route selection to an authorized owner under the current process.
“Tell me whether this design is secure.”It asks for an assessment from potentially sensitive or incomplete material.Do not enter the design. Record the unknown and request an approved assessment path.
“Generate a passing security test.”A generated result can be mistaken for verification or validation evidence.Keep only the supplied test-reference label; authorized testing happens in an approved environment.
“Mark the risk accepted.”Risk acceptance can bind owners and create legal, contractual, and operational consequences.Mark acceptance authority and status unknown; name the accountable risk owner.
“Deploy the mitigation.”Deployment is an external system change with reliability and security consequences.Stop at the review packet. Authorized humans choose and execute any change through the current process.
Runnable local artifact

Design-Assurance Evidence Desk

Enter only fictional or organization-approved, minimum-necessary labels. This form stays in the browser and creates a review-only receipt; it does not send, save, inspect, query, or connect to anything.

Primary sources

What this manual is built on.

Sources are evidence context, not permission to use private material, assess a system, choose a security action, or take an external action. Recheck them and controlling organization rules before any non-fictional use.

Related guides

Where this guide fits.

Proposed links: the Information Security Analyst Security Signal Evidence Desk for pre-investigation handoffs, the Computer Systems Engineers and Architects Evidence Desk for broader architecture decisions, the Software Quality Assurance Analysts and Testers Manual for test and release evidence, the AI Agent Workflow Builder for least-privilege job cards, and the Occupation AI Workflow Guide Directory for the broader shelf. Use these related guides to keep each workflow's authority boundary clear.