sanicode 0.11.0

AI-assisted code sanitization scanner with OWASP ASVS, NIST 800-53, and ASD STIG compliance mapping.

Sanicode

Sanicode scans code in 22 languages for input validation and sanitization gaps using field-sensitive taint analysis and a data flow knowledge graph backed by treeloom's Code Property Graph, then maps every finding to OWASP ASVS 5.0, NIST 800-53, ASD STIG v4r11, PCI DSS 4.0, FedRAMP, and CMMC 2.0. It also scans lockfiles for third-party dependency vulnerabilities via the OSV database and can generate CycloneDX 1.5 SBOMs. Output formats include SARIF (for GitHub Code Scanning), JSON, Markdown, and an HTML dashboard with an interactive knowledge graph.

Unlike pattern-only tools like Bandit or Semgrep, sanicode traces tainted data from source to sink across function boundaries with field-level precision — request.args and request.form["name"] are tracked as distinct taint keys, not flattened to request. Findings carry context about how untrusted input reaches a dangerous call and whether sanitization exists along the path.

Install

pip install sanicode

Requires Python 3.10+.

For a guided walkthrough with a sample vulnerable application, see the Getting Started Guide.

Quick start

Scan a codebase and generate a Markdown report:

sanicode scan .

Generate SARIF output for CI integration:

sanicode scan . -f sarif

Generate an HTML dashboard with an interactive knowledge graph:

sanicode scan . -f html

Generate a DISA STIG Viewer checklist for ATO packages:

sanicode scan . -f stig-checklist

Fail the build if high-severity findings exist:

sanicode scan . --fail-on high

Scan dependencies for known vulnerabilities:

sanicode deps .

Generate a CycloneDX SBOM alongside scan results:

sanicode scan . --sbom sbom.json

Reports are written to sanicode-reports/ by default.

CI/CD integration

GitHub Action

- uses: rdwj/sanicode@v0
  with:
    path: .
    fail-on: high
    format: sarif

Pre-commit hook

# .pre-commit-config.yaml
repos:
  - repo: https://github.com/rdwj/sanicode
    rev: v0.11.0
    hooks:
      - id: sanicode

See docs/ci-cd-integration.md for GitLab CI, Jenkins, Azure DevOps, and Tekton/OpenShift Pipelines.

API server

Start the FastAPI server for remote or hybrid scan mode:

sanicode serve

This starts on port 8080 with Prometheus metrics at /metrics.

Endpoints

POST /api/v1/scan              Submit a scan (async)
GET  /api/v1/scan/{id}         Poll scan status
GET  /api/v1/scan/{id}/findings   Retrieve findings (JSON or ?format=sarif)
GET  /api/v1/scan/{id}/graph      Retrieve knowledge graph
POST /api/v1/analyze           Instant snippet analysis
GET  /api/v1/compliance/map    Compliance framework lookup
GET  /api/v1/health            Liveness check
GET  /metrics                  Prometheus metrics

CLI commands

sanicode scan .                              # Scan codebase, generate reports
sanicode scan . -f sarif                     # SARIF output
sanicode scan . -f json -f sarif             # Multiple formats
sanicode scan . -f html                      # HTML dashboard with interactive graph
sanicode scan . --fail-on high               # Exit non-zero on high+ findings
sanicode serve                               # Start API server on :8080
sanicode report scan-result.json             # Re-generate reports from saved results
sanicode report scan-result.json -s high     # Filter by severity
sanicode report scan-result.json --cwe 89    # Filter by CWE
sanicode config setup                        # Interactive provider configuration wizard
sanicode config set llm.fast.model granite-nano  # Script-friendly config
sanicode config test                         # Test configured LLM tiers
sanicode config --show                       # Show resolved configuration
sanicode config --init                       # Create starter sanicode.toml
sanicode graph . --export graph.json         # Export knowledge graph
sanicode graph . --visualize graph.html      # Standalone graph visualization
sanicode rules --list                        # List all detection rules
sanicode validate-rules custom.yaml          # Validate custom rule YAML syntax
sanicode test-rules custom.yaml --fixture f.py  # Test custom rules against a fixture
sanicode benchmark                           # Benchmark against Bandit and Semgrep
sanicode scan . -f stig-checklist           # STIG Viewer checklist (.ckl) + summary
sanicode scan . -f poam                     # POA&M entries (CSV + JSON + summary)
sanicode report scan-result.json -f stig-checklist  # STIG checklist from saved results
sanicode report scan-result.json -f poam    # POA&M from saved results
sanicode enrich bandit.sarif semgrep.sarif   # Enrich third-party SARIF with compliance
sanicode enrich *.sarif --merge -o merged.sarif  # Merge and enrich multiple SARIF files
sanicode validate-llm                        # Benchmark LLM pipeline quality (precision/recall/F1 deltas)
sanicode deps .                              # Scan lockfiles for dependency vulnerabilities
sanicode deps . --format json                # JSON output for CI pipelines
sanicode deps . --sbom sbom.json             # Generate CycloneDX SBOM
sanicode scan . --no-deps                    # Skip dependency scanning
sanicode scan . --sbom sbom.json             # Include SBOM with scan
sanicode scan . --offline                    # Skip OSV queries (air-gapped mode)
sanicode scan . --no-llm                     # Deterministic-only (skip all LLM stages)
sanicode score                               # Benchmark against unsanitary-code-examples
sanicode fips-check                          # Validate FIPS 140-2/140-3 compliance
sanicode config get llm.fast.model           # Read a single config value

Detection rules

703 built-in rules across 22 languages, covering 109 CWEs including 100% of the MITRE Top 25.

Languages: Python, JavaScript/TypeScript, Go, Java, C, C++, C#, Ruby, PHP, Rust, Kotlin, Scala, Bash, SQL, Perl, Lua, MATLAB, R, F#, Julia, Fortran, COBOL.

Categories include SQL injection, OS command injection, XSS, deserialization, path traversal, SSRF, weak cryptography, hardcoded credentials, insecure random, argument injection, CRLF/header injection, XPath/LDAP/XML injection, template injection, ReDoS, XXE, mass assignment, session and cookie security, sensitive data storage, auth/authz gaps, TLS bypass, memory safety (C/C++), and many more.

For the full live inventory, see docs/coverage-scorecard.html. Custom YAML rules extend this set — place rule files in rules/ in your project root or ~/.config/sanicode/rules/ and validate with sanicode validate-rules.

Custom rules

id: CUSTOM001
cwe_id: 78
severity: high
pattern:
  targets: [python]
  ast_pattern: "call:subprocess.run"
  args:
    shell: "True"

Rule files are discovered from rules/ in the project root and ~/.config/sanicode/rules/. Run sanicode rules --validate custom.yaml to check syntax before deploying.

Taint analysis

Sanicode performs field-sensitive, dataflow-aware taint tracking at two levels:

• Intra-procedural: reaching-definitions analysis within each function body, with field-level precision. Attribute chains like request.args.get("id") are tracked as dotted taint keys, not flattened to individual identifiers. Prefix matching ensures that tainting request implicitly taints request.args, but tainting only request.args does not falsely taint unrelated attributes.
• Inter-procedural: function summaries propagated across the call graph.

Taint paths produce high-confidence edges in the knowledge graph, giving the LLM (and human reviewers) evidence of whether untrusted data actually reaches a sink. Each finding carries a node_ref field identifying its CPG node ID (e.g. "sink_3"), allowing downstream tools to traverse the graph from any finding without resolving file/line positions themselves.

Dependency scanning

Sanicode discovers lockfiles (requirements.txt, package-lock.json, composer.lock) and queries the OSV database for known vulnerabilities. Findings are mapped to CWE-1395 (Dependency on Vulnerable Third-Party Component) with compliance cross-references to NIST SI-2/RA-5, PCI DSS 6.3.2, and FedRAMP baselines. CycloneDX 1.5 SBOMs can be generated alongside scan results.

Dependency scanning runs automatically during sanicode scan and can be used standalone via sanicode deps. Use --offline for air-gapped environments or --no-deps to skip it entirely.

Compliance frameworks

Findings map to six frameworks, covering 109 CWEs:

• OWASP ASVS 5.0 — V1: Encoding and Sanitization requirements (L1/L2/L3)
• NIST 800-53 — SI-10 (Information Input Validation), SI-15 (Information Output Filtering), and related controls
• ASD STIG v4r11 — APSC-DV-002510 (CAT I), APSC-DV-002520 (CAT II), APSC-DV-002530 (CAT II), and related checks. Use --format stig-checklist to output a DISA STIG Viewer .ckl file with findings mapped directly to ASD STIG v4r11 checklist items, suitable for submission to STIG assessors.
• PCI DSS 4.0 — Requirement 6 (Develop and Maintain Secure Systems and Software)
• FedRAMP — Baselines (Low, Moderate, High) derived from NIST 800-53 control selection. Findings indicate which FedRAMP authorization baselines are affected.
• CMMC 2.0 — Cybersecurity Maturity Model Certification practices (Level 2+) mapped from NIST 800-53 controls. Useful for DoD supply chain compliance assessments.

Configuration

Create a config file:

sanicode config --init

This writes a sanicode.toml in the current directory. Config is loaded from (in order):

1. --config flag
2. sanicode.toml in the current directory
3. ~/.config/sanicode/config.toml

Sanicode works fully without any configuration. LLM tiers are optional — without them, the tool runs in degraded mode using AST pattern matching, taint analysis, knowledge graph construction, and compliance lookups. LLM integration adds context-aware reasoning on top of these. Use --no-llm to force deterministic-only mode even when LLM tiers are configured — useful for CI pipelines where speed matters more than LLM-assisted reasoning.

LLM integration (optional)

Preset-based pipeline (recommended)

The simplest way to enable LLM analysis is a single preset. Each preset selects a model, provider, and analysis strategy tuned for that model tier:

[llm]
preset = "local-medium"

Preset	Model	Strategy	F1 Score	Requirements
cloud-haiku	Claude Haiku 4.5	augment	1.000	ANTHROPIC_API_KEY
local-large	gpt-oss:20b	augment	0.970	13 GB RAM, Ollama
local-medium	granite3.3:8b	augment	0.930	5 GB RAM, Ollama
local-small	mistral-nemo	review	0.896	7 GB RAM, Ollama

Two strategies are supported. augment: the LLM analyzes code independently using CPG context, and its findings are merged with deterministic results. review: the LLM reviews deterministic findings with CPG context — better suited to mid-tier models that benefit from scaffolding. When the two perspectives disagree, a minority report is attached to the finding so both views are preserved.

Strategy guidance: Strong models perform best with augment (independent reasoning with CPG context). Mid-tier models perform best with review (reviewing deterministic findings with CPG context). Models below ~7B parameters are not recommended — accuracy drops significantly. When adding a custom model, start with augment if the model is known for strong reasoning, or review otherwise.

Legacy tiers

The three-tier system (fast / analysis / reasoning) is still supported for backward compatibility and gives fine-grained control over which model handles classification, data flow reasoning, and compliance mapping. See docs/model-sizing-guide.md for details.

Supported providers for both approaches: Anthropic, OpenAI, Google, Azure, vLLM, Ollama, OpenShift AI. Run sanicode config setup for an interactive wizard.

Current status

v0.11.0 highlights: --no-llm flag for deterministic-only scans (instant CI pipeline mode, closes #210), node_ref field on findings mapping each finding to its CPG node ID for downstream graph traversal, CWE-913 compliance enrichment fix (all five CWE-913 rules across Python/PHP/JS/Perl/Lua now get proper compliance mappings and correct derived severity), Apache-2.0 relicense, and standard contributor artifacts (CONTRIBUTING.md, SECURITY.md, PR/issue templates).

v0.10.1 — Documentation accuracy patch. Fixed stale language coverage claims in REFERENCE.md, updated llms-full.txt, regenerated coverage scorecard.

v0.10.0 highlights: CPG-backed knowledge graph via treeloom with cross-function data flow edges, augment/review LLM pipeline validated across 13 models, 703 detection rules across 22 languages (109 CWEs, 100% MITRE Top 25), CI-friendly stdout/stderr separation, FIPS compliance support, air-gapped deployment architecture, rule authoring SDK, and litellm security pin.

Plus everything from earlier releases: field-sensitive taint analysis, SBOM-aware dependency scanning via OSV, CycloneDX 1.5 SBOM generation, FedRAMP/CMMC 2.0 mappings, SARIF enrichment, POA&M generation, STIG checklist output, inter-procedural taint analysis, Grafana dashboards, MLflow integration, and CI/CD integration.

License

Apache-2.0