Releasing more often doesn't mean more secure ... or does it?

Started by Greelan, Today at 01:10:27 PM

Previous topic - Next topic
This post is written purely as a long-time OPNsense user who is interested in my own network security, and so was motivated to look into this topic. I don't purport to speak for Deciso or the OPNsense developers. This is just my take. Claude assisted with the analysis, and I had the benefit of Franco's observations.

In August a Netgate staff member posted "Releasing more often doesn't mean more secure" in the "Messages from the pfSense Team" section of their forum. The argument: release cadence measures time to patch, not defect rate. It is backed by CVE statistics: OPNsense averages 7.28 against 6.34 for pfSense, and pfSense "never once cracks 9.0".

More than a decade after the fork, Netgate still regularly measures pfSense against OPNsense in public; even a reply in that thread objected to "needlessly fanning the flames". In the same spirit, here is a look at the evidence.

What the post gets right

Cadence alone proves nothing. OPNsense did have a bad run this year: three point releases in five weeks (26.1.6 to 26.1.8), two of them for root RCEs. Every RCE cited needs an authenticated GUI user, so keeping the GUI off the WAN matters more than anyone's release schedule. And pfSense's System Patches package is a real way to ship fixes between releases.

Where the comparison falls down

1. The score depends on who does the scoring
CVEScore citedScored byVectorScored the other way
OPNsense CVE-2026-44193, -44194, -451589.1GitHubPR:H, S:C7.2, as VulnCheck scores pfSense (S:U)
OPNsense CVE-2026-57155 (GeoIP)9.9GitHubPR:L, S:C8.8 without the change of scope: the importer already runs as root. 7.2 as VulnCheck scores pfSense
OPNsense CVE-2023-39001, -39004, -39008, -271529.8NVD (the CNA gave no score)PR:NThese need authentication or local access, or aren't RCE at all (27152 is missing login rate limiting)
pfSense SA-26_23 to 26_37, 8 root RCEs (17 Nov)8.6 (CVSS 4.0)VulnCheckPR:H, no subsequent impact9.1, as GitHub scores OPNsense

There is also a structural reason scores run high. On a firewall, the web GUI is remote access by definition, and almost any configuration page can reach the OS. The page-level privilege system both projects inherited leads reporters to treat an account with access to one page as a "limited" user (PR:L), even when that page is effectively admin, and to count reaching the OS as a change of scope (S:C). All of that applies to pfSense one-for-one, but pfSense's CVEs are scored differently. VulnCheck, which issued pfSense's November CVEs, scores a GUI user as admin (PR:H) and reaching the OS as no change of scope. OPNsense's CVEs are issued through GitHub, and the cited ones use the higher reading: the three 9.1s and the 9.9 are all S:C. So the same kind of bug gets a different score depending on which project it is filed against. The DHCP hostname bug is the clearest case: 9.1 Critical for OPNsense, and no score at all for pfSense. "Never cracks 9.0" says more about who scored the bugs than about the bugs themselves.

2. CVE counts reflect disclosure channels
By Netgate's own figures, 12 of its 49 advisories since 2023 got a CVE. OPNsense opened private vulnerability reporting on GitHub in September 2025 and published its first advisory in March 2026. GitHub lets a project request a CVE for an advisory directly, and OPNsense does. Between March and the Netgate post in August, OPNsense published 28 advisories, 18 of them with CVEs. So more than half of the post's 33 OPNsense CVEs since 2023 came from five months of a new reporting channel. Before that, OPNsense's CVEs mostly came from outside researchers, often carrying NVD's worst-case default scores. Neither history makes a clean basis for averages.

3. Shared code, shared bugs, in both directions
The post cites CVE-2026-45158 (DHCP client hostname → root command execution, 9.1 Critical) against OPNsense. The code predates the fork, so both projects carried it for over a decade. OPNsense fixed it in 26.1.8 on 12 May, after an outside report. pfSense committed its fix on 29 October and released it on 17 November as SA-26_34, with no CVE and no rating.

It also went the other way. pfSense escaped LDAP usernames in July 2018 (redmine #8626) without an advisory. OPNsense carried the same bug through its rewrite of LDAP authentication until 26.1.6 in April 2026 (CVE-2026-34578, 8.2). Neither fix crossed over: pfSense's 2018 fix came with no advisory, and pfSense fixed the DHCP bug five and a half months after OPNsense's public advisory. With shared heritage, sharing reports would have fixed both bugs in both projects years earlier.

4. 17 November
pfSense released 19 advisories (SA-26_23 to 26_41): 17 with CVEs, 11 rated High. Nearly all are unsanitized GUI input reaching a shell or config file, the bug class the post itself names. Several had been in pfSense for 10 to 15 years. 11 of the 19 are in code OPNsense inherited at the 2014 fork. OPNsense had dealt with all 11 before 17 November, but not all in the same way:

Bug (pfSense SA)How and when OPNsense dealt with itAhead of pfSense
DHCP client hostname (26_34)Fixed after an outside report, 26.1.8, May 2026 (CVE-2026-45158)6 months
Time zone → php.ini (26_29)Fixed after an outside report, 26.1.11, Jun 2026 (CVE-2026-57154); a 2017 rewrite had kept the bug4.6 months
Unbound ACL name (26_40)Fixed after an outside report, 26.1.11, Jun 2026 (CVE-2026-57154); the 2023 MVC migration had kept the bug4.6 months
IPsec login banner (26_26)Quoted in 25.7 (Jul 2025); newlines blocked after an outside report, 26.1.114.6 months (partly 1.3 years)
Wireless parent interface (26_31)Shell escaping audit, 26.1, Jan 202610 months
CARP VHID (26_24)Code rewritten, 21.7.3, Sep 20215.2 years
DynDNS custom URL file read (26_33)Feature moved to a plugin, plugin removed in 23.7 (Jul 2023)3.3 years
PPPoE server username (26_38)Feature removed, 20.7 (Jul 2020)6.3 years
QinQ (26_23)Legacy feature removed, 18.7 (Jul 2018)8.3 years
PPP periodic reset (26_37)Feature removed, 16.7 (Jul 2016)10.3 years
pkg.php privilege bypass (26_27)Removed, 15.1.8 (Mar 2015)11.7 years

Six of the eleven went away because the feature was dropped, not because the bug was fixed. Fewer features means a smaller attack surface, but users lost those features. Four were fixed only in 2026, after outside reports.

5. Refactoring cuts both ways
Rewrites removed some of these bugs, but they carried others forward (time zone 2017, Unbound ACLs 2023, LDAP), and new code brings new bugs. The 9.9 GeoIP bug the Netgate post cites (CVE-2026-57155) is one; on its merits it is closer to 8.8, since the importer it exploits already runs as root. The original 2016 GeoIP code limited country codes to three characters. A 2022 change adding the EU region (23.1) dropped that limit, and 2025 support for the IPinfo format (26.1) added a second unchecked path. Both allowed path traversal until 26.1.11 (June 2026).

The longer-term work is aimed at the bug class rather than single bugs:
  • mwexec(), the unescaped shell helper inherited from pfSense, went from 86 call sites in 25.1 to one in 26.1 and is gone in 26.7.
  • A privilege-separated web GUI (not running as root) has been available as an option since 25.7 and is planned as the default in 27.1. That would change what a GUI command injection is worth.
pfSense's equivalent is the planned rewrite of its UI in Go.

So does releasing more often mean more secure?

A user's exposure has two parts: how long until a bug is found and fixed, and how long until the fix ships. Cadence clearly shortens the second part. From fix to release, OPNsense took 4 days on the DHCP bug and 18 days on the June batch; pfSense took 11 to 34 days on the November batch, with System Patches as an earlier option. Cadence plausibly shortens the first part too, because code that is reworked for a release every two weeks keeps being revisited. Of the 11 shared bugs above, 7 went away through routine work (rewrites, removal of broken or obsolete features, the shell escaping audit); none of those commits mentions a security issue, and the reasons they give are not security reports. The other 4 were fixed after reports published as GitHub advisories. The same constant change also introduces bugs, as the GeoIP case shows. Even so, most of the six-month DHCP gap had nothing to do with cadence: a bug disclosed publicly in shared code took pfSense five and a half months to pick up. The LDAP case shows the same gap the other way round.

Openness affects the first part too. OPNsense's current code is public for anyone to review. pfSense's public GitHub repository hasn't been updated since 31 March 2026, so the 17 November fixes can't be reviewed there.

Neither codebase is free of this PHP-era bug class; both carried identical bugs for over a decade. A fair comparison would score both projects on one rubric, count advisories rather than CVEs, and measure exposure windows for shared bugs. On the 19 bugs above, OPNsense users were protected earlier, sometimes by years. That came from frequent releases together with what they bring: continuous rewriting and removal of old code, work aimed at the whole bug class, and outside researchers using an open reporting channel. Frequent releases on their own would not have done it.

Method: each of the 19 bugs was traced through opnsense-core and opnsense-plugins history from the fork, and the GeoIP and LDAP bugs through their own histories. "Inherited" means the same unescaped code existed at the fork; it was not exploit-tested on OPNsense except where an OPNsense advisory confirms it. Scores come from the CVE records and NVD. Alternative scores were computed from the same vectors with the Python cvss library. Advisory counts come from the GitHub API. Release timing is measured from the commit date to the release date. Repositories checked on 21 November 2026.