neo4j-cve

Graph database version of the NVD CVE database by Jim Jazwiecki and Nathan Kluth, built using the official Neo4j Docker image (GitHub). Please post feedback/questions as GitHub issues.

Setup

Run docker-compose build to download the Neo4j Docker image and install the APOC plugin
Run docker-compose up to start Neo4j
Run docker exec -t nvdgraph python3 code/nvd_loader.py to load the NVD data feeds

Starting Neo4j

Run docker-compose up to start the container. Open http://localhost:7474/browser/ to start an interactive browser-based session. By default, there is no authentication set, but it can be set in Docker from your local environent by setting the NEO4J_AUTH environment variable to a username/password pair. For example:

export NEO4J_AUTH="admin/password"

Python Setup

There are two different Python environments used in this project. code/requirements.txt is used by the Neo4j container to load the NVD data to Neo4j, but doesn't include the Python Neo4j driver, because it is not needed by nvd_loader.py. To work with the Neo4j database locally using Python and the official Neo4j Python package, create a virtual environment and run pip install -r requirements.txt.

Python Usage

Connecting to the DB

The following will work out of the box, unless you've set the NEO4J_AUTH environment variable:

driver = GraphDatabase.driver('bolt://localhost')

Then, to create a new session:

s = driver.session()

Queries

To run a query from within a session:

s.run("MATCH n RETURN n LIMIT 25")

NVD CVE Data

Schema

Details on specific field mappings are code/loader-template.cypher, but broadly the schema is as follows:

(AttackVector) 
    -[:ATTACKABLE_THROUGH]-
(CVE)
    -[:AFFECTS]-
(ProductVersion)
    -[:VERSION_OF]-
(Product)
    -[:MADE_BY]-
(Vendor)

Loading CVE Data

https://nvd.nist.gov/vuln/data-feeds lists all feeds of data. nvd_loader.py will fetch this page, parse it to identify complete (i.e. not partial/update) v1.0 gzipped JSON feed URLs, fetch the files one by one, unpack, and load them using the template specified in loader-template.cypher.

loader-template.cypher

First, the script will set uniqueness constraints on our nodes, implicitly creating indices. $nvd_file_name will be replaced during execution with the name of the JSON file being loaded. Then we loop over all the individual CVEs.

Exploring data

Run this command to load a test file, code/nvd-test-samples.json, which contains a couple sample CVEs:

CALL apoc.load.json('file:///var/lib/neo4j/code/nvd-test-samples.json') YIELD value AS nvd
UNWIND nvd.CVE_Items as vuln
RETURN vuln

See https://neo4j-contrib.github.io/neo4j-apoc-procedures/ for more details on apoc.load.json.

Sample queries

Calculate the number of vulnerabilities found in the same software with the same impact score:

MATCH
    (vendor : Vendor {name : 'microsoft'})
        −[:MADE BY]−
    (product : Product {name: 'edge'})
        −[:VERSION OF]−
    (product version:ProductVersion)
        −[:AFFECTS]−
    (cve :CVE)
RETURN cve.`v2.impact score`, count(cve)

Calculate the number of vulnerabilities found in the same software with the same impact score across two consecutive years:

UNWIND range(1988, 2018, 1) AS t
WITH
    apoc.date.fromISO8601(t + '-01-01T00:00:00.000Z')
        AS start_window,
    apoc.date.fromISO8601((t + 1) + '-12-31T23:59:59.999Z')
        AS end_window
MATCH
    (vendor: Vendor)
        -[:MADE_BY]-
    (product:Product)
        -[:VERSION_OF]-
    (product_version:ProductVersion)
        -[:AFFECTS]-
    (cve:CVE)
WHERE
    cve.published >= start_window
        AND
    cve.published <= end_window
RETURN
    apoc.date.format(start_window)
        AS start_window,
    apoc.date.format(end_window)
        AS end_window,
    vendor.name,
    product.name,
    cve.`v2.impact_score`
        AS impact_score,
    count(cve) AS vulnerabilties
ORDER BY
    start_window,
    vendor.name,
    product.name

Find count of vulnerabilities that require user interaction through the network, under the CVSS 3.0 definition of "attack vector"

MATCH
    (attack_vector:AttackVector {name: 'NETWORK'})
        -[:ATTACKABLE_THROUGH {cvss_version: 3}]-
    (cve:CVE {`v3.user_interaction`: 'REQUIRED'})
RETURN count(cve)

Identify easy network-accessible exploits that don't involve user actions which simplify access to high-impact exploits which are otherwise high-complexity:

MATCH
    (attack_vector:AttackVector {
            name: 'NETWORK'
        }
    )
        -[:ATTACKABLE_THROUGH]-
    (cve:CVE {
        `v3.scope`: 'CHANGED',
        `v3.user_interaction`: 'NONE',
        `v3.attack_complexity`: 'LOW'
        }
    )
        -[:AFFECTS]-
    (product_version:ProductVersion)
        -[:AFFECTS]-
    (escalated_cve:CVE {
        `v3.privileges_required`: 'HIGH',
        `v3.user_interaction`: 'NONE',
        `v3.integrity_impact`: 'HIGH'}
    )
        -[:ATTACKABLE_THROUGH]-
   (escalacted_vector:AttackVector {name: 'LOCAL'})
RETURN cve.name,
    product_version.name,
    escalated_cve.name

jazwiecki/neo4j-cve