Owasp Bsi Audit

OWASP + BSI IT-Grundschutz Audit

Audits like a human auditor would — only automated and more thorough: every control assessed individually, every verdict justified. The main context stays free because the actual assessment work goes to subagents that write their results to disk (see ORCHESTRATION.md). The BSI portion follows BSI-Standard 200-2 — Schutzbedarf is fixed at normal (Basis + Standard requirements) per this skill's design.

Report language note: the generated report.md/report.html/fix-plan.md are deliberately written in German — BSI IT-Grundschutz is a German standard with German normative vocabulary (MUSS/SOLLTE, Basis-/Standard-Anforderungen, Umsetzungsstatus ja/teilweise/nein/entbehrlich), and the audience for a Grundschutz-Check is German-speaking. This skill's own documentation (this file and its siblings) is English per repo convention; only the report output is German by design.

Flow

1. Check catalog freshness. Read catalog/SOURCES.md. Older than 90 days or a file is missing → run python3 scripts/build_catalog.py (fetches the latest ASVS/MASVS/BSI/SSDF/SLSA editions, see below). Briefly tell the user if a refresh ran.
2. Confirm the target directory. If the user didn't name a path, ask or infer it from context. Create <repo>/.audit-tmp/.
3. Structure analysis. One Agent call, model: haiku, following the prompt skeleton in ORCHESTRATION.md step 1. Result: .audit-tmp/profile.json.
4. Schutzbedarf + modeling. Directly in the main thread (small enough), following BSI-METHODIK.md Phase 2+3; write .audit-tmp/methodik.json. Derive the list of candidate control groups from it.
5. Baustein/standard selection (confirm with the user). Present the candidate list from step 4 — grouped by standard (BSI Bausteine, ASVS chapters, MASVS groups if mobile targets were found, SSDF/SLSA groups if applicable) — and ask the user which ones to actually run this audit against, via AskUserQuestion (multiSelect). Default selection = everything the modeling step proposed, but the user must confirm or adjust before dispatch; never silently run the full catalog without this checkpoint. This mirrors BSI-200-2's real-world practice of a human-reviewed Modellierung. Record the confirmed selection in methodik.json (modellierung[].bausteine reflects only the confirmed set).
6. Fan-out / Grundschutz-Check. One Agent call per confirmed group, model: sonnet, following ORCHESTRATION.md step 3, in batches of 3–5 in parallel. Each writes .audit-tmp/findings/<group-id>.json and returns only a one-line summary.
7. Render. python3 scripts/render_report.py .audit-tmp/ --html (always pass --html unless the user explicitly wants Markdown only). Produces report.md, report.html, fix-plan.md — layout in REPORTING.md.
8. Wrap-up. Short summary (totals per standard, top findings) plus the three file paths, to the user. Don't paste the full findings into the chat — the files are the source of truth.

A run is done when: every confirmed control group has written a findings/<group-id>.json, render_report.py completed without error, and all three methodology-report sections (Strukturanalyse, Schutzbedarf, Grundschutz-Check) are non-empty.

Catalog scope

• ASVS 5.0 (L1+L2 = normal) — web backends, APIs.
• MASVS 2.1 — only when structure analysis finds Flutter/React Native targets.
• BSI Bausteine (Basis+Standard) — only the code-/config-checkable ones: CON.8, CON.10, APP.3.1, APP.3.2 (config-checkable parts), APP.4.3 (MySQL), APP.1.4 (mobile apps — only when mobile targets exist). Pure hosting/infrastructure requirements are tracked via out_of_scope_reason, not dropped — details in BSI-METHODIK.md. Most of the BSI Kompendium's ~111 Bausteine (ISMS, ORP, OPS, NET, INF, IND layers) are organizational-governance or physical-infrastructure and were deliberately excluded — they don't fit a code audit for a solo developer or small team without a dedicated ops/security department.
• NIST SSDF (curated, 5 of 65 practices) — only practices that are (a) not already covered by BSI CON.8/CON.10 or ASVS, (b) not the job of the separate full-quality-scan skill (static analysis, dependency scanning), and (c) code-/config-checkable without a formal organizational process. See SSDF_SELECTED_IDS in build_catalog.py.
• SLSA (curated, Build track only, 2 requirements) — supply-chain build integrity: hosted CI vs. manual builds, build provenance. The Source track is excluded (SLSA Level 4 requires two-party review, not achievable for a solo developer).

Important rules

• Every Agent call sets model: explicitly (haiku for recon, sonnet for assessment) — never inherit the default.
• BSI findings always use the vocabulary ja/teilweise/nein/entbehrlich, with a mandatory justification for entbehrlich and for nein/teilweise at Standard level (Basis requirements should be ja). OWASP/SSDF/SLSA findings use pass/fail/partial/n_a. Schema: schema/finding.schema.json.
• Read config values for real where possible (php.ini, security headers, .htaccess, TLS) — don't just assert them.
• .audit-tmp/ does not belong in the audited repo — don't commit it, just point the user to it at the end.