ca9 0.4.0

Open source Python package security and evidence-backed SCA triage

ca9

Local, evidence-backed security for Python packages and SCA alerts.

---

The problem

Your SCA tool (Snyk, Dependabot, Trivy, pip-audit, OSV, or another scanner) flags every CVE in your dependency tree. You get 60 alerts. Your team scrambles. But many of those CVEs are in code your application never imports, never calls, and never executes.

You're patching vulnerabilities in functions you don't use, in packages you didn't know you had, in code paths your app will never reach.

That's wasted engineering time. That's alert fatigue. That's how real vulnerabilities get ignored.

What ca9 does

ca9 is a local-first security layer for Python packages and open-source dependency risk. Today, its strongest path turns CVE alerts into evidence-backed fix, suppress, or investigate decisions. The newer inventory path normalizes packages, artifacts, lockfile evidence, and dependency edges so ca9 can grow beyond CVE-only reachability without making any resolver or package manager a required dependency.

It takes your CVE list and answers one question per vulnerability: is this code actually reachable from your application?

pip install ca9[cli]
ca9 scan --repo . --coverage coverage.json

For package inventory and supply-chain evidence, ca9 can read project manifests natively and uses fyn.lock or npm package-lock.json when present:

ca9 inventory --repo . -f json
ca9 vet --repo . -f json

For agentic triage, ca9 can normalize SARIF from tools such as CodeQL or Semgrep into a ca9.evidence.v1 evidence report:

ca9 ingest-sarif codeql.sarif --repo . -f json

CVE ID               Package   Severity  Verdict
--------------------------------------------------------------
GHSA-cpwx-vrp4-4pq7  Jinja2    high      REACHABLE
GHSA-frmv-pr5f-9mcr  Django    critical  UNREACHABLE (static)
GHSA-mrwq-x4v8-fh7p  Pygments  medium    UNREACHABLE (dynamic)
--------------------------------------------------------------
Total: 61  |  Reachable: 25  |  Unreachable: 36  |  Inconclusive: 0

59% of flagged CVEs are unreachable — only 25 of 61 require action

36 CVEs eliminated. No manual triage. No guessing.

How it works

ca9 combines repository evidence with advisory metadata to determine whether vulnerable code is reachable:

1. Static analysis (AST import tracing) — Parses every Python file in your repo and traces import statements. If a vulnerable package is never imported, it's unreachable.

2. Dependency inventory — Uses declared dependencies, report metadata, local package metadata, fyn.lock, and npm package-lock.json when available to separate direct, transitive, imported, and unused packages.

3. Dynamic analysis (coverage.py) — Checks whether vulnerable code was actually executed during your test suite. A package might be imported but the specific vulnerable function might never be called.

4. Advisory normalization — Preserves source, aliases, CWE/CPE IDs, timestamps, and cache freshness where input data provides them, so evidence can be traced back to the alert.

For each CVE:
  Is the affected version installed or declared?
  ├── NO  → UNREACHABLE (static)
  └── YES → Is the package, affected submodule, or known vulnerable API used?
      ├── NO, with enough graph/import evidence → UNREACHABLE (static)
      ├── YES, and runtime/coverage confirms execution → REACHABLE
      ├── YES, but coverage shows no affected execution → UNREACHABLE (dynamic)
      └── Not enough evidence → INCONCLUSIVE

ca9 is conservative — it only marks something unreachable when it can prove it. Every verdict comes with an evidence trail and a confidence score so you can see exactly why ca9 reached its conclusion.

Where ca9 fits

ca9 does not replace your SCA tool. It adds local, evidence-first reachability analysis to the vulnerability data you already have.

	ca9	Alert-only SCA output	Hosted reachability platforms
Local analysis	Runs in your repo/CI	Varies	Often requires source upload or hosted project import
Direct OSV scan	Yes — ca9 scan queries OSV.dev directly	Not always	Varies
SCA report parsing	Snyk, Dependabot, Trivy, pip-audit	Native to each tool	Platform-specific
Package inventory	Native manifests plus optional fyn.lock and npm package-lock.json artifacts and dependency edges	Varies	Varies
Static + dynamic evidence	Imports, dependency graph, coverage, API usage	Usually package-level alerts	Varies by vendor and integration
Open outputs	JSON, SARIF, OpenVEX, Markdown, HTML, remediation, action plan	Vendor-specific	Platform-specific
Confidence/evidence trail	Structured evidence per verdict	Limited	Varies
Runtime dependencies	packaging core dependency; optional CLI/MCP extras	Varies	Hosted service

Use ca9 when you want an open, local Python reachability layer for CVE triage, CI gates, SARIF upload, OpenVEX generation, or SBOM enrichment.

Real-world results

Django REST Framework — 37 CVEs, 19% noise

A focused library that genuinely uses most of its deps. Even here, ca9 found 7 CVEs in packages that are installed but never imported (redis, sentry-sdk, pip):

$ ca9 scan --repo /path/to/drf -v

GHSA-g92j-qhmh-64v2  sentry-sdk  low       UNREACHABLE (static)
                      -> 'sentry-sdk' is not imported and not a dependency of any imported package
GHSA-8fww-64cx-x8p5  redis       high      UNREACHABLE (static)
                      -> 'redis' is not imported and not a dependency of any imported package
...
Total: 37  |  Reachable: 0  |  Unreachable: 7  |  Inconclusive: 30

Flask app with bloated deps — 61 CVEs, 59% noise

A Flask app that imports 4 packages but has 19 pinned in requirements.txt (Django, tornado, Pygments added "just in case"):

$ ca9 scan --repo demo/ --coverage demo/coverage.json

Total: 61  |  Reachable: 25  |  Unreachable: 36  |  Inconclusive: 0

59% of flagged CVEs are unreachable — only 25 of 61 require action

Django alone brought 21 CVEs that were pure noise.

The pattern: ca9's value scales with how bloated your dependency list is — which in enterprise codebases is typically very.

Quick start

Scan repository dependency versions (no SCA tool needed)

pip install ca9[cli]
ca9 scan --repo .

This resolves exact dependency versions from the target repository and queries OSV.dev. ca9 does not use the ambient Python environment unless you explicitly pass --allow-env-fallback, which keeps CI scans tied to repository evidence. No Snyk, no Dependabot, no config files.

Inspect package inventory and lockfile evidence

ca9 inventory --repo . -f json

When fyn.lock or npm package-lock.json is present, ca9 reads it directly and includes package versions, direct/transitive dependency edges, dependency groups, artifact URLs, hashes or integrity values, registries, and source evidence. If there is no lockfile, ca9 falls back to native Python manifest readers for pyproject.toml, requirements*.txt, Pipfile, uv.lock, and poetry.lock.

Run supply-chain risk checks

ca9 vet --repo .
ca9 vet --repo . --malware-query
ca9 vet --repo . --scan-artifacts
ca9 vet --repo . --scan-workflows
ca9 vet --repo . --internal-package 'acme-*' --private-index https://packages.acme.internal/simple
ca9 vet --repo . --deny-license AGPL-3.0 --deny-license GPL-3.0

ca9 vet evaluates the normalized package inventory for local supply-chain risk signals: untrusted package indexes, missing artifact hashes, missing artifact metadata, source-only install risk, and mutable package sources. With --malware-query, ca9 also queries OSV for known malicious-package advisories such as MAL-*, PYSEC-MAL-*, and malware-labeled GHSA/OSV records across PyPI and npm. With --scan-workflows, ca9 checks GitHub Actions workflows for risky token scopes, OIDC write access, pull_request_target trust-boundary patterns, mutable action refs, cache trust boundaries, and source-clone commands. Direct dependencies from untrusted indexes, known malicious packages, and high-risk workflow patterns are blocking findings; weaker local signals are warnings or investigation items.

With --scan-artifacts, ca9 downloads only lockfile artifacts with hashes by default, verifies the hash, safely unpacks wheels/sdists without executing code, and runs GuardDog-style static heuristics for suspicious .pth startup execution, install-time setup.py execution, startup customization hooks, credential/network exfiltration, import-time risky behavior, silent process execution, and encoded payload execution.

Replay real incident fixtures

python scripts/incident_replay.py --strict --format table

ca9 keeps real incident fixtures for npm package compromise, PyPI import-time malware, and GitHub-token compromise scenarios. The current matrix is intentionally honest: malware advisories and workflow-risk patterns are covered where fixtures prove them, while package-tarball, import-time malware, and identity/audit-log surfaces remain partial.

For dependency-confusion controls, use --internal-package with one or more private package name patterns and --private-index for the indexes those packages are allowed to resolve from. For license policy, use --deny-license; ca9 reads wheel/sdist metadata and blocks denied direct dependencies while warning or investigating weaker cases.

Screenshot-ready supply-chain demo

Use the local demo fixture when you need a report screenshot or a JSON artifact without depending on a live suspicious repository:

bash demo/supply_chain/run_demo.sh

The fixture generates a fyn.lock with local, hash-pinned wheel artifacts and then runs ca9 vet with artifact scanning, dependency-confusion policy, and denied-license policy. The underlying gate exits 1 because the findings are intentionally blocking; the wrapper still writes demo/supply_chain/ca9-vet.json for screenshots and CI artifact examples.

ca9 supply-chain report for .../demo/supply_chain/repo
Packages: 4 | Edges: 3 | Findings: 3 | Block: 3 | Warn: 0
Artifact scans: 3 | Skipped artifacts: 0

Findings:
  [BLOCK] dependency_confusion critical acme-internal@1.0.0
    Possible dependency confusion for acme-internal
  [BLOCK] python-startup-pth-exec critical startup-hook@1.0.0
    Python startup file executes suspicious code in startup-hook
  [BLOCK] denied_license high license-risk@1.0.0
    Denied license for license-risk

Add dynamic analysis for better results

coverage run --source=.,$(python -c "import site; print(site.getsitepackages()[0])") -m pytest
coverage json -o coverage.json
ca9 scan --repo . --coverage coverage.json

Analyze an existing SCA report

ca9 check snyk.json --repo . --coverage coverage.json
ca9 check dependabot.json --repo .

Format is auto-detected. Supports Snyk, Dependabot, Trivy, and pip-audit:

ca9 check snyk.json --repo .
ca9 check dependabot.json --repo .
ca9 check trivy.json --repo .
ca9 check pip-audit.json --repo .

Verdicts

Verdict	What it means	What to do
REACHABLE	Evidence shows the vulnerable package, component, or known API is reachable	Fix this
UNREACHABLE (static)	Package is never imported — not even transitively	Suppress with confidence
UNREACHABLE (dynamic)	Package is imported but vulnerable code was never executed	Likely safe — monitor
INCONCLUSIVE	Imported but no coverage data to prove execution	Add coverage or review manually

Evidence and confidence

Every verdict is backed by structured evidence. Use --show-confidence to see scores in table output, or inspect the evidence object in JSON/SARIF output.

Signal	What it checks
advisory	Advisory source, ecosystem, aliases, CWE/CPE IDs, and cache freshness metadata when available.
version_in_range	Is the installed version within the affected range (PEP 440)?
package_imported	Is the package imported anywhere in the repo?
submodule_imported	Is the specific vulnerable submodule imported?
coverage_seen	Was the vulnerable code executed during tests?
api_call_sites_covered	Were specific vulnerable API call sites executed in tests?
coverage_completeness_pct	Overall test coverage percentage — weights dynamic absence signals
affected_component_source	How was the vulnerable component identified (commit analysis, curated mapping, regex, class scan)?

Confidence scoring is verdict-directional — evidence that supports the verdict boosts the score, evidence that contradicts it lowers it. A high confidence UNREACHABLE is different from a high confidence REACHABLE.

Bucket	Score	Meaning
High	80-100	Strong evidence supports the verdict
Medium	60-79	Moderate evidence, reasonable certainty
Low	40-59	Weak evidence, treat with caution
Weak	0-39	Very little evidence, manual review recommended

CLI reference

ca9 scan [OPTIONS]              Scan repository dependency versions via OSV.dev
ca9 check SCA_REPORT [OPTIONS]  Analyze a Snyk/Dependabot/Trivy/pip-audit report
ca9 inventory [PATH] [OPTIONS]  Show normalized package inventory
ca9 vet [PATH] [OPTIONS]        Run package supply-chain risk checks

Common options:
  -r, --repo PATH                  Path to the project repository  [default: .]
  -c, --coverage PATH              Path to coverage.json for dynamic analysis
  -f, --format [table|json|sarif|vex|remediation|action-plan|markdown|html]
                                      Output format  [default: table]
  -o, --output PATH                Write output to file instead of stdout
  -v, --verbose                    Show reasoning trace for each verdict
  --no-auto-coverage               Disable automatic coverage discovery
  --show-confidence                Show confidence score in table output
  --show-evidence-source           Show evidence extraction source in table output
  --proof-standard [strict|balanced]
                                      Proof policy for suppressions
  --capabilities                   Attach AI capability blast radius
  --runtime-context PATH           Deployment-aware severity adjustment
  --trace-paths                    Trace exploit paths
  --threat-intel                   Enrich with EPSS and CISA KEV data
  --otel-traces PATH               Production runtime evidence from OTLP JSON
  --accepted-risks PATH            Accepted-risk TOML/JSON file
  --baseline PATH                  Previous ca9 JSON report for new-only gating
  --new-only                       Only gate on new reachable/inconclusive findings

Scan-only options:
  --offline                        Use only cached OSV data, no network requests
  --refresh-cache                  Clear OSV cache before fetching
  --allow-env-fallback             Use installed package versions when repo versions cannot be resolved
  --max-osv-workers N              Max concurrent OSV detail fetches  [default: 8]

Inventory-only options:
  -f, --format [table|json]         Output format  [default: table]

Vet-only options:
  --trusted-index URL               Trusted package index; repeatable
  --private-index URL               Private index allowed for internal packages
  --internal-package PATTERN        Internal package glob, e.g. acme-*; repeatable
  --malware-query                   Query OSV for known malicious packages
  --scan-artifacts                  Hash-verify, unpack, and statically inspect artifacts
  --scan-workflows                  Scan GitHub Actions workflow risk patterns
  --allow-unhashed-downloads        Allow artifact downloads without lockfile hashes
  --max-artifact-mb N               Max artifact download size  [default: 100]
  --deny-license ID                 Denied license identifier; repeatable
  --require-known-license           Warn when artifact metadata has no known license
  --offline                         Use cached OSV data only for malware query

Exit codes:
  0  Clean — no reachable CVEs
  1  Reachable CVEs found — action needed
  2  Inconclusive only — need more coverage data

Config file

Create a .ca9.toml in your project root to set defaults:

repo = "src"
coverage = "coverage.json"
format = "json"
verbose = true
accepted_risks = "accepted-risks.toml"
baseline = "ca9-baseline.json"
new_only = true

Config is auto-discovered from the current directory upward. CLI flags override config values. Accepted-risk and baseline options keep ignored findings visible in report output while excluding them from exit-code decisions.

Caching and offline mode

ca9 caches OSV vulnerability details (~/.cache/ca9/osv/, 24h TTL) and GitHub commit file lists (~/.cache/ca9/commits/, 7-day TTL) to reduce API calls.

ca9 scan --repo . --offline           # use cached data only, no network
ca9 scan --repo . --refresh-cache     # clear cache and re-fetch

Set GITHUB_TOKEN to avoid GitHub API rate limits when ca9 fetches commit data for affected component analysis:

export GITHUB_TOKEN=ghp_...
ca9 check snyk.json --repo .

fyn integration

ca9 does not require fyn. If a repository has fyn.lock or npm package-lock.json, ca9 inventory parses it natively and treats it as high-fidelity package evidence. This gives ca9 exact resolved versions, direct/transitive edges, groups, artifact hashes or integrity values, artifact URLs, and source registries without shelling out to package manager CLIs.

ca9 vet builds on that evidence for local supply-chain checks. For example, a direct dependency resolved from a non-trusted index is treated as risky, and an internal package matching --internal-package is blocked if it resolves outside the configured --private-index values.

When --scan-artifacts is enabled, fyn's artifact URLs and hashes let ca9 verify and inspect resolved wheels/sdists before applying malicious-package heuristics. This path does not install packages or execute package code. The same artifact metadata powers license checks through --deny-license and --require-known-license.

Future ca9 commands can use fyn as an optional provider for dependency-path and lock-diff context, but absence of fyn should not break scans or CI gates.

MCP server

ca9 ships an MCP server so LLM-powered tools (Claude Code, Cursor, etc.) can run reachability analysis directly.

pip install ca9[mcp]

Add to your MCP client config:

{
  "mcpServers": {
    "ca9": {
      "command": "ca9-mcp"
    }
  }
}

Available tools:

Tool	What it does
check_reachability	Analyze an SCA report (Snyk, Dependabot, Trivy, pip-audit)
scan_dependencies	Scan repository dependency versions via OSV.dev
check_coverage_quality	Assess how reliable your coverage data is
explain_verdict	Deep-dive a specific CVE's verdict with full evidence
generate_vex	Generate OpenVEX exploitability statements
generate_remediation_plan	Generate prioritized remediation actions
scan_capabilities	Scan AI capabilities and emit an AI-BOM
check_blast_radius	Attach capability blast radius to reachable CVEs
trace_exploit_path	Trace paths to vulnerable API call sites
lookup_threat_intel	Look up EPSS and CISA KEV data
enrich_sbom	Enrich CycloneDX or SPDX SBOM JSON

Library usage

import json
from pathlib import Path
from ca9.parsers.snyk import SnykParser
from ca9.engine import analyze

data = json.loads(Path("snyk.json").read_text())
vulns = SnykParser().parse(data)

report = analyze(
    vulnerabilities=vulns,
    repo_path=Path("./my-project"),
    coverage_path=Path("coverage.json"),
)

for result in report.results:
    print(f"{result.vulnerability.id}: {result.verdict.value} (confidence: {result.confidence_score})")
    print(f"  reason: {result.reason}")
    if result.evidence:
        print(f"  source: {result.evidence.affected_component_source}")

Zero heavy dependencies

ca9's core library depends on packaging for PEP 440/version normalization and uses the Python standard library for TOML parsing on Python 3.11+. On Python 3.10, tomli is used for lockfile parsing. The click package is optional — only needed if you use the CLI. This means you can embed ca9 in CI pipelines, security toolchains, or other Python tools without pulling in a large dependency tree.

Limitations

• Static analysis traces import statements and importlib.metadata dependency trees. Dynamic imports (importlib.import_module, __import__) are not detected.
• Coverage quality directly impacts dynamic analysis. If your tests don't exercise a code path, ca9 can't detect it dynamically.
• Transitive dependency resolution requires packages to be installed. Without installed deps, ca9 falls back to direct-import-only checking.
• fyn.lock and npm package-lock.json support currently power inventory and the first ca9 vet local supply-chain checks. Optional artifact static analysis currently scans Python wheel/sdist artifacts only. Full attack detection still needs richer external intelligence for maintainer changes, release-age anomalies, typosquatting, provenance, and active malware analysis.
• Python reachability only for now; npm support is inventory evidence, not JavaScript reachability.

Development

git clone https://github.com/duriantaco/ca9.git
cd ca9
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest tests/ -v

License

MPL-2.0