Build the archlinexp.eu/www baseline from clean sources on a dedicated
site branch, per the web-baseline deploy model: vanilla Joomla 3.9.2 core
plus the developer-sourced own-code plus a vetted module layer, with the
Cadline core delta applied on the 3.9.2 base. The generator is reused
unchanged: build-baseline.sh is site-agnostic, only the content differs.
Layers:
- core/ vanilla Joomla 3.9.2 official package (installer removed)
- cadline/ own-code from the developer source (beiratkozas, maintenance,
mod_al_*, mod_alusers, mod_course_list) plus 11 assets that
the developer source lacks
- deployed/ third-party module slugs taken from live, vetted
- overrides/ the 5 genuine Cadline core modifications, on the 3.9.2 base
The live core carried 31 stale com_users files: the updater skipped the
customized ones, so the version string said 3.9.2 while the code did not.
Taking vanilla 3.9.2 plus the 2 real deltas fixes that silently.
Vetting caught two items that content-based YARA did not flag. Both come
from a structural rule: an extra file on a core-component path that is
not a template override is not a module -- it is a leftover or a plant.
Both are excluded through malware-iocs.paths:
- components/com_mailto/mail.php: an unauthenticated file-write webshell
(POST save_file + file_content -> fopen/fwrite, no auth check at all),
present on live since 2021-07-26
- administrator/components/com_joomlaupdate/restoration.php: an Akeeba
Kickstart leftover carrying a security password (known RCE vector)
Verification: build-baseline.sh -> out/ = 12284 files; YARA (16 rules) on
every source layer and on the built tree = 0 hits; none of the 291 IOC
paths present; no disguised .json dropper; core verified as 3.9.2; the
only extra file left on a core path is the legit Google reCAPTCHA library.
Assisted-by: claude-code@claude-opus-4-8
The "clean" baseline unintentionally shipped ~34 planted webshells/backdoors
plus sbtrack (a third-party SEO-cloaking proxy) and inert scanner probes,
carried in from the vetted-live Helix3 template. They survived because the IOC
safety net keyed on extension ('*.php.json') while the attacker disguised one
webshell under ~20 fake extensions (.inc.json/.phar.json/.php[3-5].json/
.php.json.json) plus plain-.json defacements.
- History rewritten (git filter-repo) to remove all 48 files from every commit.
- build-baseline.sh step 6 is now CONTENT-based: any templates/**/*.json with a
PHP open tag is dropped (legit Helix3 presets are pure JSON).
- malware-iocs.paths extended with plain-.json defacements, probe artifacts and
non-.json droppers a content grep cannot key on.
Found by the imrcli HIDS YARA union layer during baseline integration (the
baseline-diff alone would have marked them baseline-match).
Assisted-by: claude-code@claude-opus-4-8
Re-architect the baseline as constituents + a generator script (build-baseline.sh)
instead of a checked-out deployed tree: a deployed Joomla docroot differs from the
repo/source layout, so the script manifest-maps packaged extensions (lib/jmap.py),
places no-source and Cadline files, applies core overrides, excludes malware/data,
and emits out/ = the code baseline the operator HIDS verifier consumes.
Adds RUNBOOK.md (recipe+inventory), PROVENANCE.md (upstream map), malware-iocs.
Signed-off-by: LÁZÁR Imre <imre@illusion.hu>
Assisted-by: claude-code@claude-opus-4-8
