dependency-audit

name: dependency-audit description: Audit npm and Python dependencies across all Databricks Apps in a workspace. Use when checking for malicious packages, generating dependency inventories, or investigating supply chain risks. Triggers on dependency audit, npm audit, package audit, supply chain, malicious package, dependency inventory.

Disclaimer

This skill is intended to supplement your security processes and is not a substitute for security investigations. The output or skill itself may be incomplete or inaccurate for your purposes, and AI tools may hallucinate. You remain solely responsible for your security investigation and results thereof.

Auditing Databricks App Dependencies

Systematic dependency audit of all Databricks Apps deployed in a workspace. Produces a unified inventory of direct and transitive dependencies, checks for flagged/malicious packages, and identifies coverage gaps.

Prerequisites

• Databricks CLI configured with a profile for the target workspace
• Working directory for source code downloads and reports
• Sufficient workspace permissions to databricks apps list and databricks workspace export-dir

Before You Start

Environment check

Check the Python version. The system Python may be old (e.g., 3.6). Avoid features added after 3.6:

• No capture_output=True in subprocess.run — use stdout=subprocess.PIPE, stderr=subprocess.PIPE
• No f-strings if targeting 3.5 — use .format()
• No walrus operator :=

python3 --version

If you need modern Python, use uv run with inline dependencies.

Workflow

Phase 1: Enumerate Apps

List all apps in the workspace with:

databricks apps list --profile <PROFILE> -o json

Parse the JSON output. Extract:

• name
• active_deployment.source_code_path
• active_deployment.status.state
• creator

Save the apps list to apps.json for reference.

Phase 2: Download Source Code

Download all apps in parallel, for each app run:

databricks workspace export-dir <SOURCE_PATH> apps/<APP_NAME>/ --profile <PROFILE> --overwrite

Timeout handling:

• Set a 120-second timeout per app
• Log failures with the error message
• After the batch completes, retry failed downloads once with a longer timeout (300s)
• Record which apps permanently failed

Post-download validation:

• Count downloaded apps vs expected
• Print summary: succeeded, failed, timed out

Phase 3: Find All Dependency Files

Search for ALL dependency manifest and lock file types. Do not only search for package.json or requirements.txt.

# Required searches — run ALL of these:
**/package.json          # npm direct dependencies
**/package-lock.json     # npm transitive (lockfileVersion 1, 2, or 3)
**/yarn.lock             # Yarn transitive dependencies
**/bun.lock              # Bun transitive dependencies
**/bun.lockb             # Bun binary lock file (note: binary, not parseable as text)
**/requirements*.txt     # Python direct dependencies (no lock file, may include dev dependencies)
**/pyproject.toml        # Python direct dependencies (PEP 518)
**/uv.lock               # uv.lock for Python dependencies (if relevant)

For each file found, record:

• App name (first path segment under apps/)
• Relative path within the app
• Lock file type

Coverage tracking: For each app with a package.json or pyproject.toml, note whether it also has a corresponding lock file. Apps with package.json or pyproject.toml, but NO lock file are coverage gaps — their transitive dependencies cannot be confirmed. Python apps using only a requirements.txt do not have lock files, so they should be flagged as gaps as well.

Phase 4: Extract Dependencies

4a: Direct dependencies

Parse each package.json, requirements.txt, or pyproject.toml and extract direct and development dependencies. For package.json, look in dependencies and devDependencies. For Python, parse the respective formats.

For each dependency: package name, version specifier, dep type, which app, which file.

4b: Transitive dependencies from lock files

Lock files have different formats. Handle each:

package-lock.json (lockfileVersion 2 or 3):

• Parse the packages field
• Each key is a node_modules/... path — extract the package name from the last node_modules/ segment
• Extract: version, resolved URL, integrity hash

package-lock.json (lockfileVersion 1):

• Parse the dependencies field (recursive — dependencies can be nested)
• Walk the tree recursively to find all transitive deps
• Extract: version, resolved URL, integrity hash

yarn.lock:

• Text format, not JSON
• Each entry starts with a package specifier line (e.g., "react@^18.2.0":)
• Followed by indented fields: version, resolved, integrity
• Parse with line-by-line text processing

bun.lock:

• JSON-like format (Bun v1.2+ uses JSON lockfile by default)
• CRITICAL: bun.lock uses trailing commas which json.load() rejects. You MUST strip trailing commas before parsing. Use regex: re.sub(r',(\s*[}\]])', r'\1', content) to remove them.
• The packages field uses array format: each entry is "pkg-path": ["name@version", "registry-url", {metadata}, "integrity"]
• Extract package name from index 0 (strip the @version suffix), version from index 0 (after @), integrity from index 3
• Earlier versions used bun.lockb (binary) — note these as unparseable coverage gaps

uv.lock:

• Similar to bun.lock, but for Python dependencies
• Also uses trailing commas — same parsing approach as bun.lock

4c: Error handling

• Wrap each file parse in try/except
• Log malformed files but continue processing
• Never let one bad file abort the entire scan

Phase 5: Check for Flagged Packages

Search for flagged packages using three independent methods:

Method 1: Structured search — Check extracted dependency data (from Phase 4) against the flagged package list. Use case-insensitive comparison.

Method 2: Raw text search — Use Grep to search ALL files under apps/ for each flagged package name as a substring. This catches packages in vendored node_modules/, inline scripts, or non-standard locations that structured parsing would miss.

Method 3: Partial match — Search for significant substrings of the flagged names (e.g., if searching for @pypestream/floating-ui-dom, also search for pypestream and floating-ui-dom separately). Typosquat packages often use partial name matches.

Case sensitivity: ALL searches MUST be case-insensitive. EmilGroup must match emilgroup, EMILGROUP, etc.

If ANY method finds a match, flag it as a POSITIVE HIT with full context (app name, file path, matched text, which method found it).

Phase 6: Generate Reports

Produce a single unified report in both JSON and CSV formats.

JSON report (dependency_audit_report.json):

{
  "audit_metadata": {
    "workspace_profile": "...",
    "audit_date": "YYYY-MM-DD",
    "total_apps": N,
    "apps_downloaded": N,
    "apps_with_npm": N,
    "unique_packages": N,
    "total_installations": N
  },
  "flagged_package_results": {
    "packages_searched": ["..."],
    "matches_found": [],
    "verdict": "CLEAN or POSITIVE"
  },
  "coverage_gaps": [
    {"app": "...", "reason": "no lock file / download failed / binary lockb only"}
  ],
  "per_app_summary": {
    "app-name": {
      "package_files": ["..."],
      "lock_files": ["..."],
      "direct_dep_count": N,
      "transitive_dep_count": N
    }
  },
  "all_dependencies": {
    "package-name": [
      {"app": "...", "version": "...", "source": "direct|transitive", "file": "...", "resolved": "..."}
    ]
  }
}

CSV report (dependency_audit_report.csv):

app,package,version,source,dep_type,file,resolved,integrity

One row per dependency installation. Include both direct and transitive.

Phase 7: Summary Output

Print a human-readable summary:

1. Verdict on flagged packages (CLEAN or list of hits)
2. Coverage stats (apps scanned, gaps identified)
3. Top dependencies by usage count
4. Report file locations

Rationalization Table

Red Flags — STOP

If you find yourself doing any of these, you are cutting corners:

• Only searching for package-lock.json without also checking yarn.lock and bun.lock
• Skipping the raw text search because "the JSON parsing already covered it"
• Not retrying failed/timed-out downloads
• Burying coverage gaps in prose instead of listing them explicitly in the report
• Using Python features that don't work in the system Python version
• Producing separate reports instead of one unified report
• Getting 0 transitive deps from a bun.lock file (your parser is broken — check trailing commas)
• Silently catching parse errors without logging them

Examples

Example: Full workspace npm audit

User says: "Audit all Databricks Apps for malicious npm packages" Result: Enumerate all running apps, download source, scan all dependency file types, check flagged packages with 3 methods, produce unified JSON+CSV report with coverage gaps.

Example: Targeted package search

User says: "Check if any Databricks Apps use lodash.template" Result: Same workflow but with lodash.template as the flagged package. Still produce full inventory.

Example: Dependency inventory only

User says: "Give me a list of all npm packages used across our Databricks Apps" Result: Same workflow but skip Phase 5 (no flagged packages). Produce full inventory with coverage tracking.