aibom-guard

AIBOM-Guard

Generate an AI Bill of Materials, triage your EU AI Act risk tier, and turn an existing ISO 27001 ISMS into ISO 42001 readiness — from the command line, in seconds.

⚖️ AIBOM-Guard is a triage and documentation aid — not legal advice, an audit, or a conformity assessment. All output should be confirmed by a qualified human. High-risk and prohibited results must be escalated to compliance/legal.

---

Why this exists

The EU AI Act's obligations for high-risk systems become enforceable on 2 August 2026. Article 11 requires technical documentation, Article 12 requires event logging, and ISO/IEC 42001 is becoming the framework organisations use to demonstrate it. Yet:

• Commercial AI-governance platforms cost tens of thousands per year.
• Fully automated AI-BOM generation barely existed in open source.
• Most teams holding ISO 27001 have no idea what ISO 42001 adds on top.

AIBOM-Guard closes that gap for engineers: point it at a repo, get a machine-readable AI inventory, a provisional risk tier, a prioritised ISO 42001 gap list, an Annex IV draft, and an HTML compliance report — all offline, no accounts required.

---

What it does

---

Quickstart

pip install aibom-guard           # or: pip install -e ".[dev]" from source

aibom-guard all ./my-ai-project \
  --name "Hiring Assistant" \
  --use-case "resume screening and candidate ranking for recruitment" \
  --html \
  -o reports/

reports/ will contain:

---

Commands

all — full compliance pipeline

aibom-guard all ./project \
  --name "My AI System" \
  --use-case "credit scoring for loan decisions" \
  --html           # HTML dashboard
  --nist           # include NIST AI RMF crosswalk
  --llm            # Claude Haiku second opinion (requires ANTHROPIC_API_KEY)
  --validate       # exit non-zero if BOM has errors
  -o reports/

scan — AI component detection only

aibom-guard scan ./project -o reports/
aibom-guard scan ./project --format spdx -o reports/   # SPDX 3.0 output
aibom-guard scan ./project --validate                  # validate emitted BOM

classify — EU AI Act tier

aibom-guard classify ./project --use-case "medical imaging for radiology"
aibom-guard classify ./project --use-case "resume screening" --llm

crosswalk — ISO 42001 or NIST AI RMF gap analysis

aibom-guard crosswalk                          # ISO 42001 (default)
aibom-guard crosswalk --no-iso27001           # without ISO 27001 baseline
aibom-guard crosswalk --framework nist        # NIST AI RMF 1.0

validate — BOM validation

aibom-guard validate reports/aibom.cdx.json
aibom-guard validate reports/aibom.cdx.json --full   # full JSON Schema check

collect — evidence collectors

aibom-guard collect ./project \
  --github myorg/my-repo \   # maps branch protection / SAST / secret scanning to ISO 42001
  -o reports/

---

Sample output

Running against the bundled demo app (resume screening + facial recognition) correctly flags it as high-risk on two Annex III grounds:

EU AI Act tier (provisional)       🔴 HIGH-RISK
AI components detected             10
ISO 42001 readiness (vs 27001)     35%
ISO 42001 net-new controls         13

Matched categories:
  [high]    A3-1-biometrics  — biometric, facial recognition, face-recognition
  [high]    A3-4-employment  — resume screening, candidate ranking, hiring
  [limited] T1-chatbot       — openai, anthropic

---

How it works

               ┌──────────────┐
  target repo ─►    scanner   ├─► AIComponent[]  (160+ sigs, 13 API patterns, HF model IDs)
               └──────┬───────┘
                      ▼
       ┌──────────────┼───────────────┬─────────────────┬───────────────┐
       ▼              ▼               ▼                 ▼               ▼
  ┌─────────┐  ┌──────────┐  ┌──────────────┐  ┌───────────┐  ┌──────────────┐
  │ AI-BOM  │  │ EU AI    │  │ ISO 27001 →  │  │ NIST AI   │  │ Annex IV     │
  │CDX+SPDX │  │ Act tier │  │ 42001 gaps   │  │ RMF cross │  │ docgen       │
  └─────────┘  └────┬─────┘  └──────────────┘  └───────────┘  └──────────────┘
                    │ (optional)
                    ▼
             LLM second opinion
             (Claude Haiku, never
              downgrades tier)
                    │
                    └───────────────────────────────────────────┐
                                                                ▼
                                                   compliance_report.html / .md

The accuracy lives in four editable YAML knowledge bases under src/aibom_guard/data/:

Improving coverage usually means editing YAML, not code.

---

MCP server

Install the MCP server to call AIBOM-Guard directly from Claude Desktop or Cursor:

pip install "aibom-guard[mcp]"

Add to ~/.claude_desktop_config.json (or equivalent):

{
  "mcpServers": {
    "aibom-guard": {
      "command": "aibom-guard-mcp"
    }
  }
}

Available MCP tools: scan, classify_risk, iso_gaps, nist_rmf, validate, full_report.

---

Optional dependencies

pip install "aibom-guard[validate]"   # full JSON Schema BOM validation
pip install "aibom-guard[llm]"        # LLM-assisted classification (Claude Haiku)
pip install "aibom-guard[mcp]"        # MCP server for Claude Desktop / Cursor
pip install "aibom-guard[all]"        # everything above

---

Benchmark results

Five real open-source AI repos scanned — see docs/benchmarks.md.

* Tier after false-positive keyword fix (see benchmarks doc for methodology). MONAI correctly triggers HIGH-RISK on the A3-medical Annex III category.

---

Built to pair with Claude Code

This repo ships .claude/skills/ (four custom skills) and a compliance-reviewer subagent. Install ECC to get the full harness.

# Run AIBOM-Guard as MCP tools directly from Claude
aibom-guard-mcp

# Or drive the CLI from Claude Code
claude --dangerously-skip-permissions -p \
  'aibom-guard all . --name "MyApp" --use-case "hiring AI" --html -o reports/'

---

Standards referenced

• Regulation (EU) 2024/1689 (EU AI Act) — Articles 5, 6, 11, 12, 50, 72; Annex III; Annex IV
• ISO/IEC 42001:2023 — AI management systems (full 38-control Annex A)
• ISO/IEC 27001:2022 — information security management (crosswalk baseline)
• NIST AI RMF 1.0 — GOVERN / MAP / MEASURE / MANAGE (65 subcategories)
• CycloneDX 1.6 — AI-BOM serialisation format
• SPDX 3.0 — AI Profile

---

License

MIT — see LICENSE.

---

Triage aid only. Not legal advice. Confirm all findings with qualified human experts.