Agentic Security Review Skill

Purpose

Create practical security review artifacts for CompleteTech LLC agentic development workflows. Use this skill before launch, before granting new access, after material configuration changes, and after incidents or near misses.

System Boundary

This skill owns security, safety, permissions, data, credential, tool, and launch-risk review. Use it alongside discovery, proposal, or delivery when risk needs a dedicated artifact. It does not replace agentic-delivery-skill launch checklists, agentic-contract-skill legal terms, external compliance certification, formal penetration testing, or counsel-reviewed privacy/security advice.

Core Workflow

1. Identify the review event: launch, new tool, sensitive data, external action, retrieval/RAG, credential change, dependency change, incident, or signoff.
2. Gather verified facts: workflow purpose, users, data classes, tools, permissions, credentials, integrations, retrieval sources, human approvals, logs, deployment path, rollback owner, incident contacts, and known constraints.
3. Use references/use-case-decision-table.md to choose the right review artifact.
4. Use references/security-positioning.md for CompleteTech LLC security language and guardrails.
5. Use references/security-catalog.md for the artifact library.
6. Keep the review bounded and evidence-based. Do not claim compliance, certification, legal approval, penetration-test completion, production readiness, or guaranteed security unless the user provides verified evidence.

Artifact Selection Guide

• Starting a new agentic workflow: use agentic-risk-intake.
• Adding a tool or integration: use tool-permission-inventory.
• Handling API keys, tokens, service accounts, or secrets: use credential-secret-handling-checklist.
• Accessing sensitive, client, personal, regulated, or proprietary data: use data-exposure-review.
• Testing prompt injection or tool misuse: use prompt-injection-test-plan.
• Adding retrieval/RAG, indexed docs, websites, or knowledge stores: use retrieval-source-trust-review.
• Reviewing human-in-the-loop controls: use approval-gate-audit.
• Sending emails, creating calendar events, modifying files, posting messages, purchasing, billing, or changing production systems: use external-action-review.
• Needing traceability, audit trails, or operational evidence: use logging-auditability-review.
• Changing model, provider, system prompt, tool runtime, or safety settings: use model-provider-configuration-review.
• Defining storage, deletion, or retention behavior: use data-retention-review.
• Adding packages, services, scripts, or vendor dependencies: use dependency-supply-chain-review.
• Reducing access scope or sandboxing execution: use sandbox-least-privilege-checklist.
• Preparing for launch: use production-readiness-security-checklist.
• Deciding what blocks launch: use launch-blocker-checklist.
• Preparing a backout path: use rollback-plan.
• Responding to a security event or near miss: use incident-response-plan.
• Defining who to contact and when to escalate: use human-escalation-procedure.
• Summarizing adversarial testing: use red-team-test-report.
• Recording final approval status: use security-signoff-memo.

When several artifacts fit, start with the artifact closest to the change or decision being reviewed, then add supporting artifacts only when they materially reduce risk.

Quality Rules

• Use verified contact routing. Do not invent client, security, legal, billing, support, or approval email addresses; ask for the right address or use TBD.
• Preserve least privilege: name each tool, permission, credential, data class, and external action that is actually needed.
• Protect human approval gates for irreversible actions, client-facing communication, payments, data export/deletion, production changes, and material business decisions.
• Separate facts from recommendations. Label unknowns, assumptions, residual risks, blockers, and owner decisions.
• Recommend technical escalation when secrets may be exposed, logs are missing, sandboxing is weak, prompt injection can trigger tools, approval gates are bypassed, production rollback is unclear, or sensitive data flows are not understood.
• Recommend client or human approval before launch, before expanding permissions, before connecting client systems, before sending external communications, and before closing incident follow-up.

Resource Guide

• references/security-positioning.md: load for CompleteTech LLC review language and boundaries.
• references/use-case-decision-table.md: load when choosing a security review artifact.
• references/security-lifecycle.md: load for review flow from intake through launch and post-incident follow-up.
• references/security-catalog.md: load for the reusable artifact templates.
• references/template-index.json: machine-readable artifact metadata used by the renderer.
• scripts/render_security_review.py: list security artifacts or render a draft with placeholders.

Renderer

python3 scripts/render_security_review.py --list
python3 scripts/render_security_review.py --stage launch --list
python3 scripts/render_security_review.py --template agentic-risk-intake --var client_name=Acme --var workflow="support triage agent"

Rendered artifacts are drafts. Replace placeholders with verified project facts before sending, storing, or relying on them.

Rendering to a Branded PDF

Artifacts from this skill are delivered as branded CompleteTech LLC PDF documents, not raw Markdown. The renderer emits the PDF (and prints the Markdown) in one command, using the same reportlab branding engine as the contract skill:

pip install -r requirements.txt
python3 scripts/render_security_review.py --template security-signoff-memo \
  --out artifact.pdf --png artifact.png \
  --title "Security Signoff Memo" --doc-type "SECURITY REVIEW" \
  --subtitle "Workflow: <b>Support Email Triage Agent (Pilot)</b>" --meta "MEMO NO.=SEC-2026-0090" --meta "DATE=2026-06-17" \
  --var client_name="Client Name" --var workflow="support triage"

• --no-pdf emits Markdown only (the original behavior); --no-cover drops the cover page.
• Already drafted the Markdown yourself? Render it directly: python3 scripts/render_pdf.py --markdown artifact.md --out artifact.pdf --logo assets/logo.png --title "...".
• The PDF supports a Markdown subset: #/##/### headings, paragraphs, - bullets, tables, > callouts, **bold**, and [PAGE_BREAK]. PDF requires reportlab; the optional --png preview requires pypdfium2 and pillow. See assets/examples/ for a rendered example.

Network Boundary

This skill is local-only. It does not include outbound network helpers, callbacks, or any helper that posts security-review run metadata to an external service.