fleet/docs/Using-Fleet/Vulnerability-Processing.md

Vulnerability processing

• What to expect
• Coverage
• Configuration
• Performance
• Detection pipeline

What to expect

Vulnerability processing in Fleet detects vulnerable software installed on your hosts. To see what software vulnerability processing covers, check out the Coverage section on the Vulnerability processing doc.

Vulnerable software can have one or more vulnerabilities (CVEs).

For Fleet Premium users, each CVE includes its CVSS base score (reported by the National Vulnerability Database), probability of exploit (reported by FIRST), and whether or not their is a known exploit in the wild (reported by the Cybersecurity & Infrastructure Security Agency).

Fleet's strategy for detecting vulnerabilities (CVEs) varies according to the host's platform. For macOS and Windows hosts, CVEs are detected using the National Vulnerability Database (NVD). For Linux hosts, CVEs are detected using the official OVAL definitions maintained by the different publishers (Canonical, Red Hat etc.).

Windows/MacOS hosts

First, Fleet retrieves the installed software for each host using osquery queries. Then, Fleet translates each installed software into Common Platform Enumeration (CPE) names.

Then, using the CPEs, Fleet searches the list of Common Vulnerabilities and Exposure (CVE) identifiers listed in the NVD to detect the CVEs that match the defined CPEs.

If matches are found, they are exposed on each host's Host details page and on the Home page in the Fleet UI. The CVEs are also exposed in the fleetctl get software command and the GET api/v1/fleet/hosts/{id} and GET api/v1/fleet/software API routes.

Vulnerability processing happens on the Fleet instance and not on the host machine. Because of this, detected vulnerabilities cannot be used in the same way you would use an osquery query (e.g. you wouldn't be able write a query to retrieve all CVEs).

Linux hosts

First, we determine what Linux distributions are part of your fleet (keep in mind that there will be a small delay between the time a new Linux hosts is added and the time the host is 'detected'). We then use that information to determine what OVAL definitions need to be downloaded and parsed - you can find a list of all the OVAL definitions we use here. OVAL definitions will be refreshed on a daily basis.

Finally, we look at the software inventory of each host and execute the assertions contained in the corresponding OVAL file - any match is reported using the same channels as with Windows/Mac OS vulnerabilities

Coverage

For Windows/Mac OS Fleet attempts to detect vulnerabilities for installed software that falls into the following categories (types):

Apps

• macOS
• Windows

Browser plugins

• macOS, Windows
  • Chrome extensions
  • Firefox extensions

Packages

• macOS

  • Python
  • Packages installed using Homebrew

• Windows

  • Python
  • Atom
  • Packages installed using Chocolatey

For Linux, we adhere to whatever is defined in the OVAL definitions, except for:

• Kernel vulnerabilities.
• Vulnerabilities involving configuration files.

As of right now, the following distributions are supported:

• Ubuntu
• RHEL based distros (Red Hat, CentOS, Fedora, and Amazon Linux)

As of right now, only app names with all ASCII characters are supported. Apps with names featuring non-ASCII characters, such as Cyrillic, will not generate matches.

The ingestion of software varies per platform. For each platform, we run an osquery query to ingest software.

Configuration

When upgrading to Fleet 4.7.0 or later, vulnerability processing is automatically enabled if vulnerability processing and software inventory are not explicitly disabled.

If you explicitly disabled vulnerability processing, and now would like to enable this feature, first enable the software inventory feature by setting the following app config:

---
apiVersion: v1
kind: config
spec:
  features:
    enable_software_inventory: true

Then, enable vulnerability processing by specifying a path where Fleet will download the different data feeds. This can be done by setting the following app config:

---
apiVersion: v1
kind: config
spec:
  vulnerability_settings:
    databases_path: /some/path

Or through environment variables:

FLEET_VULNERABILITIES_DATABASES_PATH=/some/path

The path specified needs to exist and Fleet needs to be able to read and write to and from it. This is the only mandatory configuration needed for vulnerability processing to work. Additional options, like vulnerability check frequency, can be found in the configuration documentation.

You'll need to restart the Fleet instances after changing these settings.

Performance

Windows/Mac OS

Vulnerability processing is performed in one Fleet instance. If your Fleet deployment uses multiple instances, only one will be doing the work.

In order to conduct vulnerability processing, Fleet downloads the following files:

1. A preprocessed CPE database generated by FleetDM to speed up the translation process: https://github.com/fleetdm/nvd/releases
2. The historical data for all CVEs and how to match to a CPE: from https://nvd.nist.gov/vuln/data-feeds

The database generated in step 1 is processed from the original official CPE dictionary https://nvd.nist.gov/products/cpe. This CPE dictionary is typically updated once a day.

The matching occurs server-side to make the processing as fast as possible, but the whole process is both CPU and memory intensive.

For example, when running a development instance of Fleet on an Apple Macbook Pro with 16 cores, matching 200,000 CPEs against the CVE database will take around 10 seconds and consume about 3GBs of RAM.

The CPU and memory usages are in burst once every hour (or the configured periodicity) on the instance that does the processing. RAM spikes are expected to not exceed the 2GBs.

Linux

As with Windows/Mac OS, vulnerability detection for Linux is performed in a single Fleet instance. The files downloaded will vary depending on what distributions are on your fleet. The list of all the OVAL files we use can be found here.

When determining what specific file(s) to download we use the reported OS version and map that to an entry in the oval_sources.json dictionary. The mapping rules we use are fairly simple, depending on the distribution, we either use the major and minor versions and the platform name (for example Ubuntu 22.4.0 -> ubuntu_2204) or just the major version (for example Red Hat Enterprise Linux 9.0.0 -> rhel_09).

To reduce memory footprint during the evaluation phase and because of performance reasons, all downloaded OVAL files are parsed, and the result is stored in a file following the following naming convention: fleet_oval_platform_date.json.

The performance will be a function of three variables:

• The size of the OVAL file
• The amount of hosts to scan
• The amount of installed software

That said, the performance characteristic should be linear (if scanning 200 hosts take ~20 seconds, then scanning 2000 hosts should take ~200 seconds).

Detection pipeline

There are several steps that go into the vulnerability detection process. In this section we'll dive into what they are and how it works.

The process has different parts that are more error-prone than others. Each OS and each application developer and maintainer can (and do) have their own way of defining each part of their app. Some Linux distributions are very strict, but each distribution handles things differently.

The whole pipeline exists to compensate for these differences, and it can be divided in two sections:

1. Collection:

graph TD;
    host1[Host1 send software list]-->normalize[Normalization of names, versions, etc]
    host2[Host2 send software list]-->normalize
    host3[Host3 send software list]-->normalize
    normalize-->store[Storage for later processing]

2. Processing

Processing happens in a loop and varies depending on the platform - first Windows/Mac OS hosts will be processed, then we look at Linux hosts. The default interval is 1hr.

General process

graph TD;
    interval{Once an hour}-->normalize[Normalized software list]
    normalize-->process1[Process Windows/Mac OS hosts]
    process1-->process2[Process Linux hosts]
    process2-->interval

Windows/Mac OS

graph TD;
  process[Process Windows/Mac OS hosts] -->downloadCPE(Download CPE database from Fleet)
  downloadCPE-->cpeTranslate[CPE translation]
  cpeTranslate-->cveDownload(CVE datastreams downloaded)
  cveDownload-->cveMap[CVE detection]

Linux

graph TD;
  process[Process Linux hosts] --> fresh{OVAL defs older than one day?}
  fresh --no--> execute(Analyze hosts using OVAL definitions)
  fresh --yes--> remove(Remove old OVAL definitions)
  remove --> download(Download new OVAL definitions)
  download --> parse(Parse OVAL definitions)
  parse --> execute

Ingesting software lists from hosts

The ingestion of software varies per platform. We run a UNION of several queries in each:

• macOS
• Windows
• Linux

This is the first step into normalizing data across platforms, as we try to get all the same data for all different types of software we detect vulnerabilities on.

Ingestion can be resource hungry, both on the hosts and the Fleet server. A lot of work has gone into reducing the resources needed, and it's still ongoing.

Translating to CPE

With a somewhat normalized list of software, in order to search CVEs for it, we need to derive a CPE from the vendor, name, version, and OS.

As described briefly above, we do this by translating the NVD database of CPEs into a sqlite database that helps Fleet do the lookup of CPEs very quickly.

How accurate is this translation process?

This is the most error prone part of the process. The CPE can have some vagueness. This means that parts of it can be a *, which means when you match that CPE to a CVE it can match any of that part of the CPE.

If the CPE is too vague, the extreme case being all parts are *, all CVEs will match. You want a very specific CPE, but not too specific that a small error would make it not match a CVE (false negative).

Let's look into some examples of this stage.

Example: tmux

tmux is a Unix terminal utility to multiplex ttys. It appears listed like this in macOS:

osquery> SELECT * FROM homebrew_packages WHERE name='tmux';
+------+----------------------------+---------+
| name | path                       | version |
+------+----------------------------+---------+
| tmux | /opt/homebrew/Cellar/tmux/ | 3.2a    |
+------+----------------------------+---------+

If we look at the official releases the version we get is the same as the one listed. This means that it'll be easy to map it to a CPE that will accurately represent the software.

Now let's look at Chrome on macOS:

osquery> select name, bundle_version from apps where name like '%Chrome%';
+-------------------+----------------+
| name              | bundle_version |
+-------------------+----------------+
| Google Chrome.app | 4758.102       |
+-------------------+----------------+

Now things start to get slightly more tricky. We have to remove the .app suffix from the name, then derive the first word as the vendor and the second as the app name. We could use bundle_name for the app name, but nothing stops the app developer of adding the vendor to bundle_name, so a similar parsing would have to happen.

These are two illustrative examples. The reality is that there is no map or list of all the software available and how it's presented in each platform, so the "software to CPE" translation process is going to be evolving constantly.

Improving accuracy

In order to improve the accuracy of matching software to CPEs, CPE translations rules are added for known cases where matching fails. server/vulnerabilities/cpe_translations.json contains these rules and is included in the NVD release.

Example: ruby@2.7 installed via homebrew

The following CPE translation rule is used to reduce false positives when ruby is installed via homebrew. This is needed because ruby is commonly included in the title in the CPE database. This rule matches the software name ruby matching a regular expression pattern and installed using homebrew. When searching for CPEs, the specifed product and vendor will be added to the filter critera.

[
  {
    "software": {
      "name": ["/^ruby(@.*)?$/"],
      "source": ["homebrew_packages"]
    },
    "translation": {
      "product": ["ruby"],
      "vendor": ["ruby-lang"]
    }
  }
]

CPE Translations (array[CPE Translation Entry])

CPE Translation Entry (object)

The CPE translation rule.

Name	Type	Description
software	array[CPE Translation Software]	The CPE translation software match criteria.
translation	array[CPE Translation]	The CPE translation.

CPE Translation Software (object)

The CPE translation software match criteria. Used to match software collected from hosts. Fields are are AND'd together. Values inside each field are OR'd together.

Name	Type	Description
name	array[string]	The software name to match. Enclose within / to specify a regular expression pattern.
bundle_identifer	array[string]	The software bundle identifier (MacOS apps only) to match. Enclose within / to specify a regular expression pattern.
source	array[string]	The software source to match. Enclose within / to specify a regular expression pattern.

CPE Translation (object)

The CPE translation. Used to match CPEs in the CPE database. Fields are are AND'd together. Values inside each field are OR'd together.

Name	Type	Description
product	array[string]	The CPE product.
vendor	array[string]	The CPE vendor.
target_sw	array[string]	The CPE target software.

Matching a CPE to a CVE

Once we have a good CPE, we can match it against the CVE database. We download the data streams locally and match each CPE to the whole list. The matching is done using the nvdtools implementation.