Apache hardening: block .git and .env disclosure

Preventive, not incident-driven. No LUPMIS host is known to have been
compromised; these rules exist so the most common automated attack against a
PHP deployment cannot succeed.

Scanners probe every public host for /.env (database passwords, API keys) and
/.git/. The second is the more damaging: given /.git/index together with
/.git/objects an attacker can reconstruct the whole repository, including its
history — so any credential ever committed is exposed even after rotation.
That is directly relevant here, because the minio-uploads integration has had
access keys written into its PHP files and those files are tracked in Gitea.

- public/.htaccess: deny hidden paths and files that should never be served.
- docs/apache-hardening.conf: the same rules for vhost/server config, which is
  where they belong — .htaccess depends on AllowOverride. Also covers logging
  the real client address through the openresty proxy (mod_remoteip), and
  blocking at the proxy so requests never reach the application.

Two details that are easy to get wrong, and are handled:

- <FilesMatch "^\."> does NOT stop /.git/config. FilesMatch tests the basename
  only, and there the basename is "config". Blocking a hidden directory needs a
  rule that sees the whole path: mod_rewrite in .htaccess, <DirectoryMatch> in
  server config.
- The SPA fallback only rewrites paths that do not exist (!-f), so a real .env
  on disk skips it and is served as a plain file. The deny rules therefore run
  before the fallback, not after.

.well-known is exempt, or ACME certificate renewal would silently break.

Verified against a live Apache instance with a planted .env and .git tree:
/.env and /./.env return 403; /.git/config, /.git/HEAD, /.git/objects/... and
/backup.sql return 404; /.well-known/acme-challenge/..., /assets/*.js and
/manifest.json still return 200. Nothing in the document root is caught — the
only dotfile there is .htaccess, which Apache never serves.

These rules are damage limitation. The fix is to keep .git and .env out of a
document root: deploy build output rather than a working copy, and hold secrets
in environment variables outside the served tree.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
ekke 2026-08-11 12:53:20 +02:00
parent 7d26c740df
commit 664c242980
3 changed files with 240 additions and 0 deletions

49
dist/.htaccess vendored
View File

@ -13,9 +13,58 @@ DirectoryIndex index.php index.html
SetHandler application/x-httpd-php
</FilesMatch>
# Second line of defence for hidden FILES (.env, .htpasswd, …), in case
# mod_rewrite is not loaded. This does not cover hidden directories such as
# /.git/config — see the mod_rewrite block below, which does.
# Both Apache 2.4 and 2.2 syntaxes are given so this works either way.
#
# Access control runs before mod_rewrite, so a hidden file answers 403 here
# while a hidden directory answers 404 from the rewrite below. Both refuse to
# serve the content, which is the point; the differing codes are cosmetic.
<FilesMatch "^\.">
<IfModule mod_authz_core.c>
Require all denied
</IfModule>
<IfModule !mod_authz_core.c>
Order allow,deny
Deny from all
</IfModule>
</FilesMatch>
<IfModule mod_rewrite.c>
RewriteEngine On
# ------------------------------------------------------------------
# Block hidden paths (.git, .env, .svn, …) — MUST stay first
# ------------------------------------------------------------------
# Scanners routinely fetch /.env and /.git/config, and if /.git/index
# plus /.git/objects/… can be read the whole repository — including its
# history and any credential ever committed — can be reconstructed.
#
# Two things make this easy to get wrong:
#
# 1. <FilesMatch "^\."> does NOT stop /.git/config. FilesMatch tests
# the basename only, and the basename there is "config". Blocking a
# hidden *directory* needs a rule that sees the whole path, which in
# .htaccess means mod_rewrite (<DirectoryMatch> is server-config only).
#
# 2. The SPA fallback further down only rewrites paths that do NOT
# exist (!-f / !-d). A real .env on disk therefore skips the
# fallback and gets served as a plain file. So this rule has to come
# before it, not after.
#
# .well-known is deliberately exempt: blocking it breaks Let's Encrypt
# (ACME) certificate issuance and renewal.
#
# 404 rather than 403, so a probe learns nothing about what exists.
RewriteCond %{REQUEST_URI} !^/\.well-known/
RewriteRule (^|/)\. - [R=404,L]
# Files that are never meant to be fetched over HTTP, if one is ever
# deployed into the document root by mistake.
RewriteRule \.(sql|sqlite|sqlite3|db|bak|backup|old|orig|save|swp|log|ini|conf|key|pem|crt|p12|pfx)$ - [R=404,L,NC]
RewriteRule (^|/)(composer\.(json|lock)|package(-lock)?\.json|yarn\.lock|Dockerfile|docker-compose\.ya?ml|\.dockerignore)$ - [R=404,L,NC]
# Clean URL for the iframe embed endpoint: /embed → embed.php
# Must come BEFORE the SPA fallback so /embed doesn't get routed to
# index.php. Query strings (?mode=permit&...) pass through automatically.

142
docs/apache-hardening.conf Normal file
View File

@ -0,0 +1,142 @@
# ============================================================================
# LUPMIS2 — Apache hardening against source-code and secret disclosure
# ============================================================================
#
# WHY THIS EXISTS
# ---------------
# This is preventive hardening. No LUPMIS host is known to have been
# compromised — these rules are here so that the most common automated attack
# against a PHP deployment cannot succeed.
#
# Scanners continuously probe every public host for two things:
#
# /.env — application secrets: database passwords, API keys
# /.git/… — the repository, if a working copy was deployed
#
# The second is the more damaging of the two. Given /.git/index together with
# /.git/objects, an attacker can reconstruct the entire repository: every file
# and every past commit. That means any credential ever committed is exposed,
# even if it has since been changed, because it remains in the history.
#
# That last point is directly relevant to LUPMIS. The minio-uploads
# integration has had access keys written into upload.php, download.php and
# get_file_url.php, and those files are tracked in Gitea — so the keys are in
# the history regardless of later rotation. The exposure is only theoretical
# while that repository is private and its .git directory is not web-served;
# these rules keep the second half of that sentence true.
#
# WHICH HOSTS
# -----------
# Apply to EVERY vhost. The risk is not equal across them:
#
# • The PWA is not currently exposed — its document root is dist/, which
# contains neither .env nor .git.
# • Any app deployed by copying a working directory (a git clone) into the
# document root is exposed by construction. The minio-uploads project, for
# example, carries a .git directory and composer files.
#
# Checking is quick, from outside the network:
# curl -sS -o /dev/null -w '%{http_code}\n' https://HOST/.env
# curl -sS -o /dev/null -w '%{http_code}\n' https://HOST/.git/config
# Anything other than 403 or 404 needs attention.
#
# HOW TO USE
# ----------
# Preferred — in the vhost or server config, where it cannot be disabled by a
# stray AllowOverride and applies to the whole tree:
#
# <VirtualHost *:443>
# ...
# Include /etc/apache2/conf-available/lupmis-hardening.conf
# </VirtualHost>
#
# Then: a2enconf lupmis-hardening && apachectl configtest && systemctl reload apache2
#
# If you can only use .htaccess, the equivalent rules are already in the PWA's
# public/.htaccess — but note .htaccess needs AllowOverride to be enabled, so
# the vhost form is the safer of the two.
#
# THIS IS DEFENCE IN DEPTH, NOT THE FIX
# -------------------------------------
# The real fix is that .git and .env must not be inside a document root at all.
# Deploy build output (or `git archive`), not a working copy, and keep secrets
# in environment variables outside the served tree. These rules are what
# protects you when that goes wrong again.
# ============================================================================
# ---------------------------------------------------------------------------
# 1. Hidden directories and files (.git, .env, .svn, .hg, .DS_Store, …)
# ---------------------------------------------------------------------------
# <DirectoryMatch> is server-config only, which is exactly the context this
# file is meant for, and it covers the whole path — so unlike <FilesMatch> it
# does stop /.git/config.
#
# .well-known is exempt on purpose: blocking it breaks Let's Encrypt (ACME)
# certificate issuance and renewal.
<DirectoryMatch "/\.(?!well-known)">
Require all denied
</DirectoryMatch>
<FilesMatch "^\.">
Require all denied
</FilesMatch>
# mod_rewrite equivalent — also covers odd forms such as /./.env, which the
# captured log shows being used, and returns 404 rather than 403 so a probe
# learns nothing about what is present.
<IfModule mod_rewrite.c>
RewriteEngine On
RewriteCond %{REQUEST_URI} !^/\.well-known/
RewriteRule (^|/)\. - [R=404,L]
</IfModule>
# ---------------------------------------------------------------------------
# 2. Files that should never be served, if one lands in the document root
# ---------------------------------------------------------------------------
<FilesMatch "\.(sql|sqlite|sqlite3|db|bak|backup|old|orig|save|swp|swo|log|ini|conf|key|pem|crt|p12|pfx|yml|yaml)$">
Require all denied
</FilesMatch>
<FilesMatch "^(composer\.(json|lock)|package(-lock)?\.json|yarn\.lock|Dockerfile|docker-compose\.ya?ml|Makefile|README\.md|\.htpasswd)$">
Require all denied
</FilesMatch>
# ---------------------------------------------------------------------------
# 3. Do not advertise the software version, and do not list directories
# ---------------------------------------------------------------------------
# ServerTokens / ServerSignature are global directives — put them in the main
# server config, not inside a <VirtualHost>.
#
# ServerTokens Prod
# ServerSignature Off
Options -Indexes
# ---------------------------------------------------------------------------
# 4. Proxy — make the logs usable
# ---------------------------------------------------------------------------
# Where Apache sits behind the openresty reverse proxy, it logs the proxy's
# Docker-internal address rather than the real caller. If that is the case
# here, an access log cannot tell you who probed the site, and the traffic
# cannot be blocked or reported. Have the proxy pass the caller through:
#
# proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
# proxy_set_header X-Real-IP $remote_addr;
#
# and have Apache trust it, so %a in the log is the real address:
#
# RemoteIPHeader X-Forwarded-For
# RemoteIPInternalProxy 172.20.0.0/16
# LogFormat "%a %l %u %t \"%r\" %>s %b \"%{Referer}i\" \"%{User-Agent}i\"" combined
#
# (RemoteIPHeader needs mod_remoteip: a2enmod remoteip)
#
# Blocking at the proxy as well as here is worthwhile — it stops the request
# before it reaches the application at all:
#
# location ~ /\.(?!well-known) { return 404; }

View File

@ -13,9 +13,58 @@ DirectoryIndex index.php index.html
SetHandler application/x-httpd-php
</FilesMatch>
# Second line of defence for hidden FILES (.env, .htpasswd, …), in case
# mod_rewrite is not loaded. This does not cover hidden directories such as
# /.git/config — see the mod_rewrite block below, which does.
# Both Apache 2.4 and 2.2 syntaxes are given so this works either way.
#
# Access control runs before mod_rewrite, so a hidden file answers 403 here
# while a hidden directory answers 404 from the rewrite below. Both refuse to
# serve the content, which is the point; the differing codes are cosmetic.
<FilesMatch "^\.">
<IfModule mod_authz_core.c>
Require all denied
</IfModule>
<IfModule !mod_authz_core.c>
Order allow,deny
Deny from all
</IfModule>
</FilesMatch>
<IfModule mod_rewrite.c>
RewriteEngine On
# ------------------------------------------------------------------
# Block hidden paths (.git, .env, .svn, …) — MUST stay first
# ------------------------------------------------------------------
# Scanners routinely fetch /.env and /.git/config, and if /.git/index
# plus /.git/objects/… can be read the whole repository — including its
# history and any credential ever committed — can be reconstructed.
#
# Two things make this easy to get wrong:
#
# 1. <FilesMatch "^\."> does NOT stop /.git/config. FilesMatch tests
# the basename only, and the basename there is "config". Blocking a
# hidden *directory* needs a rule that sees the whole path, which in
# .htaccess means mod_rewrite (<DirectoryMatch> is server-config only).
#
# 2. The SPA fallback further down only rewrites paths that do NOT
# exist (!-f / !-d). A real .env on disk therefore skips the
# fallback and gets served as a plain file. So this rule has to come
# before it, not after.
#
# .well-known is deliberately exempt: blocking it breaks Let's Encrypt
# (ACME) certificate issuance and renewal.
#
# 404 rather than 403, so a probe learns nothing about what exists.
RewriteCond %{REQUEST_URI} !^/\.well-known/
RewriteRule (^|/)\. - [R=404,L]
# Files that are never meant to be fetched over HTTP, if one is ever
# deployed into the document root by mistake.
RewriteRule \.(sql|sqlite|sqlite3|db|bak|backup|old|orig|save|swp|log|ini|conf|key|pem|crt|p12|pfx)$ - [R=404,L,NC]
RewriteRule (^|/)(composer\.(json|lock)|package(-lock)?\.json|yarn\.lock|Dockerfile|docker-compose\.ya?ml|\.dockerignore)$ - [R=404,L,NC]
# Clean URL for the iframe embed endpoint: /embed → embed.php
# Must come BEFORE the SPA fallback so /embed doesn't get routed to
# index.php. Query strings (?mode=permit&...) pass through automatically.