vulnometry

Measure what a vulnerability is worth to your business, not how severe it is in the abstract.

The problem

Your scanner tells you a CVE exists and that NVD rated it 9.8. It does not know that the affected library sits on a decommissioned lab box, or that the unexciting 6.1 is on the payment gateway and in PCI scope. So everything critical looks equally urgent, and the queue gets worked in the wrong order.

Vulnometry takes the same public data your scanner uses, combines it with a description of what you actually run, and gives you one answer per place you run the affected thing.

$ vulnometry compare CVE-2021-44228

CVE-2021-44228: one CVE, 3 answers
Asset            BEI  Verdict     Threat  Reach  Conseq.  Due
checkout-api   928.9  Contain       1.00   0.90     1.00  2026-09-04
internal-wiki  164.5  Schedule      1.00   0.16     0.41  2027-01-06
ml-sandbox       0.0  Accept        1.00   0.00     0.22  -

Threat is identical everywhere. The 929-point spread is entirely reachability
and consequence. That difference is what knowing your own estate buys you.

Same CVE, three jobs, three dates, three owners.

Contents

• Install
• Five-minute tour
• Reading the output
• How the score works
• Describing what you run
• Bulk analysis
• Dashboard
• Commands
• Using a model
• Other surfaces
• Resilience
• Limits

Install

Requires Python 3.10 or newer.

pip install vulnometry

The optional extras are only needed for specific backends:

pip install 'vulnometry[bedrock]'    # + AWS Bedrock
pip install 'vulnometry[service]'    # + the REST server
pip install 'vulnometry[all]'        # both

To work on the code instead, clone it and use the install script, which sets up a virtualenv and runs the tests:

git clone https://github.com/san3ncrypt3d/vulnometry.git
cd vulnometry
./install.sh                    # Windows: .\install.ps1
source .venv/bin/activate       # Windows: .venv\Scripts\activate

No API keys are required for anything. Two free ones make it considerably faster, and vulnometry doctor will tell you if you need them:

export NVD_API_KEY=...    # nvd.nist.gov/developers/request-an-api-key, raises 5 req/30s to 50
export GITHUB_TOKEN=...   # any GitHub token, no scopes needed, for advisory lookups

Five-minute tour

vulnometry doctor                    # check the feeds are reachable
vulnometry inventory init            # describe what you run, in one YAML file
vulnometry compare CVE-2021-44228    # one CVE, one answer per asset

To try it against a worked example estate before describing your own:

curl -O https://raw.githubusercontent.com/san3ncrypt3d/vulnometry/main/examples/vulnometry.yaml
vulnometry compare CVE-2021-44228

Then point it at a real scanner export:

vulnometry import ~/Downloads/scan.xlsx

That reads your scanner export, scores every row against your inventory, and writes an annotated workbook plus a self-contained HTML dashboard into the current directory.

Step-by-step version with explanations: QUICKSTART.md.

Reading the output

Every row of every report has the same shape.

BEI is the three factors multiplied together, so a zero in any one of them takes the whole score to zero. That is deliberate: something nobody can reach is not urgent, however frightening its CVSS score.

Due dates come from policy rather than the score alone: tier 1 assets halve the window, regulated systems get 60% of it, production 80%.

Every assessment carries the reasoning that produced it, so you can see which feed contributed what. Run vulnometry measure CVE-2021-44228 --asset checkout-api for the long form.

How the score works

BEI = 1000 * Threat^0.8 * Reachability^0.7 * Consequence^0.6

Three questions, multiplied rather than summed:

Multiplication is the design decision. A weighted sum lets a frightening CVSS carry a finding nobody can reach, which is how teams end up patching lab machines while something dull sits on the gateway. Multiplied, any factor can veto. That falls out of the arithmetic, so there is no override branch in the code and nothing to argue about in a review meeting.

The cost is that unknowns matter. If you have not described an asset, vulnometry treats the gaps as uncertainty rather than safety, and scores it pessimistically. That is on purpose: losing a finding is worse than over-ranking one.

The full derivation, including every weight and exponent and the reasoning behind it, is in docs/SCORING.md. Disagreements about the weights are welcome and make good issues.

Describing what you run

This is the part that makes the tool useful, and it is one YAML file. A minimal one:

assets:
  - name: checkout-api
    tier: 1
    internet_exposed: true
    data_classification: restricted
    regimes: [pci-dss]
    owner: payments@example.com
    components: ["org.apache.logging.log4j*"]

Only name is required. Every other field sharpens the measurement:

Leave a field out rather than guessing. Unknown is treated as uncertainty, which is safer than a confident wrong answer.

vulnometry inventory init writes a commented template to start from, and vulnometry inventory show prints what it parsed. Vulnometry finds the file automatically if it is named vulnometry.yaml and sits in the working directory, or you can pass --inventory.

Bulk analysis

Most vulnerability work arrives as a spreadsheet somebody was emailed.

vulnometry import nessus-export.xlsx --preview      # check column detection first
vulnometry import nessus-export.xlsx                # measure everything

Always run --preview first. It shows which columns were recognised, without spending any API calls, so you find out immediately if the CVE column was missed.

Columns are detected by header against an alias table, and the header row is found rather than assumed, because exports routinely carry a title block above the table. A row listing several CVEs becomes several findings. Assets are matched to your inventory by name, hostname glob or component glob.

Reads .xlsx, .xls, .csv and .tsv, plus native JSON from Trivy, Grype, Snyk, OSV-Scanner, Dependabot and SARIF. vulnometry sweep handles lockfiles and SBOMs: requirements.txt, poetry.lock, package-lock.json, go.mod, Gemfile.lock, Cargo.lock, pom.xml, CycloneDX and SPDX.

The workbook it writes has four sheets: Findings (every row, colour-coded, filterable, frozen header), Action Plan (only what needs doing, sorted by due date), By Owner (workload per team), and Method (how the numbers were produced).

Dashboard

vulnometry import export.xlsx --dashboard exposure.html

One self-contained HTML file. No server, no CDN, no build step and no JavaScript framework; charts are inline SVG generated in Python. It opens from file:// and works air-gapped, which matters when the person who needs the summary is not the person with a terminal.

It shows a KPI strip, a verdict donut, exposure by business unit, actionable load by owner, and a reachability-versus-consequence scatter with bubble size set by threat. Findings in the top-right corner are the ones to work on.

Also served at /dashboard when running vulnometry serve http.

Commands

measure and import take a --lens controlling how much evidence to gather:

• signal (default) uses the CVE record, EPSS and KEV. Fast enough for thousands of rows.
• full adds OSV and GHSA, so the answer includes a fix version.
• forensic adds a search for public exploit code.

Both also take --fail-on contain|remediate|schedule, which exits non-zero when anything reaches that verdict. That is the hook for CI; see examples/ci-workflow.yml.

Run vulnometry <command> --help for the full set of options.

Using a model

The deterministic path never calls a model, so measure, import and compare work on an air-gapped laptop with no inference stack installed. A model is only ever used to explain a verdict, never to produce one.

When you do want narration, one action registry feeds every provider:

vulnometry ask "what must ship before Friday?" --model ollama:qwen3
vulnometry ask "..." --model openai:gpt-4o-mini
vulnometry ask "..." --model anthropic:claude-sonnet-4-6
vulnometry ask "..." --model bedrock:us.anthropic.claude-sonnet-4-20250514-v1:0
vulnometry ask "..." --model azure:my-gpt4o-deployment       # Foundry and Azure OpenAI
vulnometry ask "..." --model onprem:Qwen2.5-72B-Instruct     # vLLM, llama.cpp, LM Studio, TGI
vulnometry ask "..." --model gemini:gemini-2.0-flash

vulnometry providers            # what is configured
vulnometry providers --probe    # what actually answers

Other surfaces

vulnometry serve mcp          # Claude Desktop, Claude Code, Cursor, Zed, Cline, Goose
vulnometry serve http         # REST + /openapi.json for Codex, custom GPTs, n8n
vulnometry actions --dialect openai   # schemas for your own agent loop

As a library:

from vulnometry import assess_finding
from vulnometry.inventory import AssetProfile

result = await assess_finding("CVE-2021-44228",
    asset=AssetProfile(name="checkout-api", tier=1, internet_exposed=True))
print(result.exposure.verdict, result.exposure.index, result.due_by)

See examples/ for working scripts, including bring_your_own_agent.py if you already have an agent loop and just want the schemas.

Resilience

Triage tooling that dies when one upstream is down fails at the moment you need it most.

• Feeds are fetched concurrently and failures are isolated. A dead NVD costs you CVSS, not the report.
• NVD falls back to the CVE Program CNA record automatically, and records which one it used.
• Rate limits are enforced client-side with per-host token buckets, so you throttle yourself instead of collecting 429s.
• Responses cache to SQLite. The KEV catalogue is one 2 MB document and a 4,000-row import fetches it once. Without the cache, bulk analysis is not practical at polite request rates.
• Action errors come back as data rather than exceptions.
• Missing evidence lowers confidence and is listed in gaps instead of being rounded away.

Limits

Worth knowing before you rely on it:

• Reachability is inventory-derived. Nothing here verifies that the vulnerable code path is actually invoked in your deployment; that is what commercial reachability analysis does.
• Exploit-artifact discovery is a heuristic GitHub search. A match means public code claims to exploit the CVE, not that it works.
• EPSS is a daily prediction. A missing score usually means the CVE is very new, not that it is safe.
• The tier, data-class and regime weights are defensible defaults, not empirical constants. Read the Method sheet and edit them to match your organisation.
• Unknown context is treated as uncertainty rather than safety, so vulnometry will over-score an asset you have not described. That is deliberate.

Prior art

Combining NVD, EPSS and KEV for prioritisation is well-trodden ground: CVE_Prioritizer, OSV-Scanner, Grype and several MCP servers all work this territory. What is different here is the business inventory as a first-class scoring input, the multiplicative model that lets context veto, and the bulk spreadsheet workflow that most security teams actually live in.

Contributing

Scoring disputes are the most useful issues this project gets; the weights are reasoned defaults, not empirical constants. See CONTRIBUTING.md for setup and the two rules that keep the codebase in shape.

Security issues: please use private vulnerability reporting rather than a public issue. See SECURITY.md.

License

Apache-2.0. See LICENSE and NOTICE.