Security findings: unsafe reflection, Jinja2 without autoescaping, unpinned GitHub Actions

[DataWizual]

During a security audit of napalm, three findings were identified.

---

Finding 1: Unsafe reflection via importlib.import_module() (High — CWE-470)

importlib.import_module() is called with a dynamic value in two locations:

• napalm/base/__init__.py:88 — core library
• napalm/eos/eos.py:174 — Arista EOS driver

If user-controlled input reaches this call — via device hostname, config file, or API parameter — an attacker can load arbitrary code. In a network automation context this runs with access to device credentials and configuration state.

---

Finding 2: Jinja2 template rendering without autoescaping (High — CWE-79)

napalm/base/helpers.py:112 uses Jinja2 without autoescape=True:

env = jinja2.Environment(loader=jinja2.FileSystemLoader(...))

This is production code, not docs. Three detectors flagged it independently (Semgrep + Bandit B701). If template data originates from device responses, injection is reachable.

Fix: jinja2.Environment(loader=..., autoescape=True)

---

Finding 3: Unpinned GitHub Actions in CI/CD workflows (Medium — CWE-1357)

Both pythonpublish.yml and commit.yaml use tag-pinned actions instead of SHA-pinned:

- uses: actions/checkout@v3

The publish workflow has access to PyPI credentials. A compromised upstream action tag silently injects into the release pipeline.

Fix: Pin to full commit SHA, e.g. actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683

[ktbyers]

@bewing On that eos.py import, is there any reason we need to support/pass a string here (as opposed to just pass a subclass of EapiConnection (or pass a string of one of the pyeapi.client.TRANSPORTS[transport])?

Is anyone actually using this.

Obviously, not super important, but on its face seems somewhat unnecessary?

[ktbyers]

Finding #2 looks to be N/A as we are not using a web-browser and not rendering HTML (though you might render XML). Even XML probably doesn't matter in the context we use it.

Also you should control your templates so it is on the NAPALM user not to allow situations where malicious input could be used.

Also there is a very good chance turning on auto-escaping would break things for NAPALM's template use.

[ktbyers]

Finding #3 improved/fixed here:

#2324