ai-security-audit

AI Security Audit Skill v2.0

Overview

This skill conducts a comprehensive AI security and privacy audit. It maps every AI touchpoint, researches actual vendor policies, scores risk weighted by data sensitivity, and produces a full output package: interactive scorecard, prioritized action plan, financial exposure estimates, data flow diagram, exportable policy templates, safer alternatives for high-risk tools, and a drift monitoring checklist.

Mode Selection

Choose the mode based on user intent:

Mode A: Pre-Push Security Gate (default for GitHub push requests)

Use this mode when the user says things like "push to github", "git push", "open a PR and push", or asks for a pre-push security check.

Goal: fast go/no-go decision before code publication.

Use only these references by default:

• references/categories.md
• references/evaluation-domains.md
• references/regulatory-profiles.md (only if regulated data/jurisdiction risk appears)
• references/alternatives-database.md (only if a finding is Orange/Red)

Do not generate full audit artifacts in this mode unless explicitly requested. Skip scorecard UI, policy templates, drift checklist, and full action-plan package.

Pre-Push Workflow

1. Scope the push surface
   • Identify what is being pushed (repo/app/service), and whether changes involve code assistants, LLM APIs, CI/CD AI, meeting bots, customer data, secrets, or regulated data.
2. Run targeted intake
   • Use the categories.md model but only for directly relevant tools and integrations in this push path.
3. Run focused scoring
   • Score at least Domains 2, 3, 4, 5, 6.
   • Include Domains 1 and 7 when data flow or operational continuity risk is implicated.
4. Return gate decision
   • PASS: no critical blockers detected.
   • PASS WITH CONDITIONS: push allowed only if listed fixes are completed immediately.
   • BLOCK: do not push until blockers are resolved.

Pre-Push Output Format

## Pre-Push AI Security Gate

Decision: PASS | PASS WITH CONDITIONS | BLOCK

Scope reviewed:
- ...

Top findings:
- [Severity] [Finding] -> [Why it matters]

Required before push:
- [Concrete fix]

Recommended after push:
- [Follow-up improvement]

Mode B: Full AI Security Audit

Use this mode for full organizational audits and governance outputs.

Workflow - Full Audit Mode (6 Phases)

Phase 1: Context — Regulatory Profile & Organization Shape

Before mapping tools, establish context that affects scoring weight. Read references/regulatory-profiles.md for question set and weight mappings.

Ask upfront (use ask_user_input_v0):

1. Where does the organization operate? (US / US+EU / Global / Other)
2. What regulations apply? (General / HIPAA / Financial / Government / Education / None specific)
3. Process EU resident data? (Yes / No / Unsure)
4. Most sensitive data category? (Investor materials / Client PII / Health / Financial / Code-IP / Strategic plans / None)
5. Team size? (1-5 / 6-20 / 21-100 / 100+)

This drives Domain 6 (Regulatory) calibration and data sensitivity weighting.

Phase 2: Intake — Map Surface with Sensitivity & Access

Walk through tool categories per references/categories.md. For each tool, capture:

1. Tier (Free / Paid / Enterprise)
2. Data sensitivity — classify per the 4-tier model:
   • 🔴 Tier 1 Critical: investor materials, M&A, legal/privileged, board, trade secrets, credentials
   • 🟠 Tier 2 Sensitive: client PII, financials, strategic plans, proprietary code, HR, health data
   • 🟡 Tier 3 Internal: comms, project plans, drafts, non-privileged meeting notes, design
   • 🟢 Tier 4 Low: public content, marketing, published docs, general research
3. Who uses it (Everyone / Team / 1-2 people / Unknown)
4. Approval status (Approved / Organic adoption / Unknown)

Present full inventory with sensitivity and access columns for confirmation.

Phase 3: Research — Investigate + Identify Alternatives

Research per references/evaluation-domains.md. Enhancements:

• For any tool trending Orange or Red during research, immediately look up alternatives in references/alternatives-database.md and quick-score them
• Track vendor policy last-updated dates — flag changes within 12 months
• Note whether protections are contractual (MSA/DPA) or policy-only

Phase 4: Score — Weighted Evaluation + Financial Exposure

Score per references/evaluation-domains.md with sensitivity weighting:

• Tier 1 data: Domains 2, 5, 6 at 2.0x weight
• Tier 2 data: Domains 2, 4, 6 at 1.5x weight
• Tier 3-4 data: All domains at 1.0x

Assess blast radius per tool (Org-wide / Team / Individual / Unknown→Org-wide).

For Orange and Red findings, estimate financial exposure per references/cost-model.md.

Phase 5: Output — Generate the Full Package

5 outputs (save all to /mnt/user-data/outputs/):

1. Interactive Scorecard (React .jsx) — per references/scorecard-template.md. Includes sensitivity badges, weighted scores, financial exposure, alternatives for Orange/Red tools.

2. Quick Wins Action Plan (.md) — per references/action-plan-template.md. Four time horizons: Day 1 / Week 1 / Month 1 / Quarter 1. Standalone doc for ops handoff. This is the most important output.

3. Detailed Report (.md) — Full findings with regulatory context, data sensitivity table, weighted scores, financial exposure, alternatives, policy references, monitoring checklist.

4. Data Flow Diagram — Use Figma generate_diagram (if connected) or Visualizer to show: tools → providers → retention periods → contractual protections. Makes the architecture legible at a glance.

5. Policy Templates (.md files) — Tailored from references/policy-templates.md:

   • AI Acceptable Use Policy (pre-filled with findings)
   • Vendor Risk Questionnaire (based on 7 domains)
   • Meeting AI Policy (if notetakers found)

Phase 6: Monitor — Drift Detection Setup

Generate a Monitoring Checklist per references/drift-monitoring.md:

• Vendor trust/security URLs to check quarterly
• Settings to verify monthly (Privacy Mode toggles, training opt-outs)
• Regulatory calendar (upcoming effective dates)
• Shadow AI discovery prompts
• Re-audit trigger events
• Recommended cadence based on risk profile

---

Reference Files

Read these as needed during each phase:

• references/categories.md — Tool categories, intake questions, data sensitivity classification
• references/evaluation-domains.md — 7 domains, scoring criteria, weighted scoring math
• references/regulatory-profiles.md — Regulatory intake, jurisdiction-specific weight adjustments
• references/alternatives-database.md — Safer alternatives lookup for common tools
• references/cost-model.md — Financial exposure estimation methodology
• references/policy-templates.md — AI AUP, vendor questionnaire, meeting AI policy templates
• references/action-plan-template.md — Quick Wins output format
• references/scorecard-template.md — React scorecard output specs
• references/drift-monitoring.md — Ongoing monitoring checklist format

Key Guidelines

• Data sensitivity changes everything. A 3.5 tool processing investor decks is more dangerous than a 2.5 processing marketing copy. Always classify before scoring.
• Alternatives must be verified. Quick-score any suggested alternative before recommending it.
• Financial estimates: conservative, sourced, ranges. Use IBM/Ponemon baselines. Never invent numbers.
• Policy templates must be tailored. Fill with actual findings, tool names, and configurations from this audit.
• Quick Wins is the most important output. If the user reads nothing else, this should be enough to act on.
• The data flow diagram tells the story better than any score table for board/investor audiences.
• Tier matters enormously. Same tool, different tier = different risk profile. Always verify.
• Shadow AI is the biggest gap. Push gently. Most orgs have 3-5x more AI than they think.