Skip to content

Netdryrun 0.4.1: the false-pass release

A verifier that says "passed" without really checking is worse than no verifier. 0.4.1 is about that one thing.

  • Fixes the false passes in our two earlier advisories: the record prefix range (GHSA-cg9f-xjw5-38ph) and the 23 cases in GHSA-g4f6-4624-2pjv. One part of case 19 stays a documented limit on SR Linux and Junos. Each fixed case has a test that must fail.
  • Publishes four narrower ones that our final red team found on the release candidate, with a workaround for each: on Arista EOS, SR Linux EVPN, FRR/SONiC VRF pings, and SR Linux record --attributes (a regression in 0.4.1). They are fixed first in 0.4.2. Known false passes in 0.4.1.
  • 0.3.3 and 0.4.0 are yanked from PyPI. Upgrade with pipx upgrade netdryrun and pin netdryrun/Netdryrun@v0.4.1.
  • Record your baselines again. 0.4.1 refuses some files 0.4.0 accepted, such as a prefix without a length or an IPv6 bgp_prefixes neighbor without afi. Every such refusal is a check that could never fail before.
  • Also new: Cisco NX-OS and MikroTik RouterOS drivers (beta, used only when you choose them), the capture kit (netdryrun capture and sanitize), and contributor gates on every pull request.

How it was tested: static checks, the unit suite, mutation testing, both verdict corpora, live labs on every example across 7 network OSes, our end-to-end new-user test, and a red team on the final candidate. One layer, the scripted end-user walk-through, was not run on the final commit; the release review says why.

Found after release by the same new-user test on a fresh GitHub runner (both fail safe: an error, never a wrong pass; fixes in 0.4.2):

  • SR Linux on containerlab 0.83: netdryrun run examples/srl-frr stops with exit 2, "runs in the namespace of … which is not a lab node". The check that keeps a run from testing the host refuses SR Linux's network namespace there. We're confirming which containerlab versions are affected; the FRR examples run normally.
  • FRR 10.2 with record: it can't read one of FRR's replies, exits 1 and lists each guard it couldn't write. Until the fix, add those guards by hand.

Think you can make it say "passed" when it shouldn't? Open a corpus case and it becomes a permanent must-fail test.

Full changelog · Release on GitHub