8 Commits

Include renames

Author	SHA1	Message	Date
Thomas Vincent	41780277d3	fix(docker): restrict development permissions and start the poller (#8453) ... * fix(docker): restrict the dev stack to localhost and make the poller run The web port bound to every interface, named volumes hid the shipped .htaccess denies, and the cron job could neither find php nor write to PID 1's output, so the poller never ran. Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * test(docker): keep harnesses polling and restore the developer's config.php check_docker.sh deleted an existing include/config.php on exit, health polling aborted under pipefail before the container reported, the MariaDB wait had no bound, and the BSD mktemp template left its Xs literal. Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * docs(changelog): add fragment for #8453 Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * test(e2e): move the pointer off the Deepness multiselect button Deepness sets a zero body margin, so the unhovered #ms-button sits at the viewport origin and headless Chromium on Linux intermittently applies :hover to it. Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * ci: pull pinned integration database from authenticated mirror Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> --------- Signed-off-by: Thomas Vincent <thomasvincent@gmail.com>	2026-10-23 08:17:48 -04:00
Thomas Vincent	e2a5f32ee9	fix(ci): [develop] harden dependency installs and Docker build boundaries (#8294) ... * security(ci): constrain dependency installs and container build boundaries Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * fix(docker): retain shipped runtime libraries in build context Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * fix(tooling): quote gettext paths and honor Docker base selection Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * test(auth): use certificate expiry truth in rendering assertions Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * fix(tooling): finish Docker layer and regex scan findings Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * fix(ci): preserve documentation badge anchors without alternation Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * test(ci): preserve contributor badges and enforce Python coverage Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * ci: share changed-code coverage contracts and enforce shell policy Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * fix(ci): enforce private review files and run path regressions Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * test: measure changed gettext shell coverage Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * test: pin coverage to the reviewed PR revision Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> --------- Signed-off-by: Thomas Vincent <thomasvincent@gmail.com>	2026-10-14 22:52:37 -04:00
3 people	1e5329dd89	security: consolidate verified GHSA-rjvj win32 cmd.exe hardening for develop (#8056) ... * fix(win32): harden the cmd.exe shell sinks on develop (GHSA-rjvj-r52f-8v5q) Forward-port of the 1.2.x win32 command-injection remediation. On Windows the SNMP/ping/data-query/trap commands run through cmd.exe, which ignores the \" escape and toggles quoting on every ", so cacti_escapeshellarg cannot stop & | ^ < > ( ). - add cacti_escapeshellarg_cmd(), stripping those operators on win32 before quoting (a Unix pass-through) - snmp_escape_string(): strip the operators in the win32 branch - route the SNMP binary hostname, the data-query device values (get_script_query_path), the data-input arg builders, the Windows ping hostname, and the SNMP trap notification arguments through the strip get_script_query_path is device-reachable (a rogue monitored device supplies an SNMP index/system field that executes on the Windows poller); the rest are post-auth. Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * fix(win32): strip %VAR% from the hostname on Windows (GHSA-rjvj-r52f-8v5q) cmd.exe expands %VAR% even inside quotes, so a device hostname of %USERNAME% would expand into the SNMP/ping command line. Add an opt-in $strip_env flag to cacti_escapeshellarg_cmd() and set it for the SNMP binary and ping hostnames, which never legitimately contain a percent. Community and credential values are left untouched because they can legitimately contain %. Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * docs(changelog): note win32 cmd.exe hardening for develop * security: require Package Management permission to trust new package signers package_accept_key() (reachable via the repository-based package import flow) inserted new keys into package_public_keys, the global trust store consulted by import_validate_signature() for every future import by every user, without checking for the Package Management (realm 29) permission that gates package_keys.php/package.php/package_repos.php for the same table. An Import Templates (realm 17) user could therefore add themselves as a trusted signer for the whole install. Also add the 'accept' action to the CSRF-protected bad_actions list so the acceptance step requires POST plus a valid token like every other state-changing action. GHSA-gxgm-8rhv-r7jv * docs(changelog): note package-signer trust authorization fix * security: close tab-based open redirect bypass in CSRF timeout handler sanitize_uri()'s guard (20298c6cd) models the WHATWG rule that browsers strip leading C0 controls/whitespace before resolving a redirect, but Chromium separately strips an embedded tab from anywhere in the target. A request target of /%09/evil.example passes the guard (no leading control char, not two leading slashes) but Chromium reduces the resulting 'Location: /<TAB>/evil.example' header to a network-path reference and navigates off-site. 1.2.x already uses validate_redirect_url() for this same sink (3338a63e1), which resolves the URL via parse_url() and checks host/port against the server-configured values, so it isn't bypassable this way. Switch develop's CSRF timeout handler to match. GHSA-c725-4jm5-9fx5 * security: authorize single-item Report actions against ownership reports.php sets \ = true, so include/auth.php grants an implicit guest session (when the Guest User feature is enabled) without ever evaluating the realm 22 gate declared for this page. The single-item action handlers reports_item_remove()/_moveup()/_movedown() relied solely on that page-level gate and performed no independent ownership check, so an unauthenticated guest-bound request could delete or reorder reports_items rows belonging to any report, including ones owned by an administrator. reports_form_actions() already gained an equivalent \ check; add the same authorization (Reports admin realm 21, or the requesting user owns the item's report) to the three single-item handlers. Same underlying chain as the 1.2.31 fix for reports_user.php (guest bind removed + cacti_authorize_resource() added to reports_item_remove()); the rewritten develop reports.php/lib/html_reports.php did not carry either half of that fix forward. GHSA-j5jh-fhwg-r56x * security: authorize tree node CRUD actions against per-tree access tree.php dispatches create_node/delete_node/move_node/rename_node/copy_node directly to api_tree_*() with no is_tree_allowed() check, even though the normal tree-view path already enforces that boundary. A non-admin user with the Trees realm but no explicit grant on a given tree_id could still create, delete, move, rename or copy nodes on that tree by calling these actions directly, bypassing the authorization boundary enforced elsewhere. GHSA-prxj-84qq-v5jq * security: remove shell from Realtime RRD create and validate DS bounds poller_realtime.php built the Realtime-graph RRD 'create' command line (including the SNMP-sourced ifSpeed/ifHighSpeed interface speed, substituted in verbatim as the DS maximum) and ran it through shell_exec(). A hostile or spoofed monitored device could return an ifSpeed value such as '1000000000\uid=197608(ladams) gid=197121 groups=197121' and get arbitrary OS command execution as the web/poller user the next time any user with the Realtime Graphs permission (or an unauthenticated guest, where configured) viewed a realtime graph for an affected data source. - lib/rrd.php: rrdtool_function_interface_speed() now casts ifHighSpeed/ ifSpeed to int before they become the DS maximum. - lib/rrd.php: rrdtool_function_create() now rejects any non-numeric rrd_minimum/rrd_maximum right before they are concatenated onto the rrdtool command line, forcing them to 'U' and logging instead. - poller_realtime.php: the RRD create is now routed through the same shell-free rrdtool_execute()/proc_open(bypass_shell) path every other RRDtool invocation already uses, matching the pattern 1.2.x already uses for this exact call site, instead of shell_exec(). GHSA-f88m-q789-g22j * security: authorize Report save/edit against report ownership reports_form_save()'s update path fetched the existing report's owner from the database but never compared it to the session user before calling sql_save(), so any user holding only the Reports realm could overwrite another user's report by POSTing its id. reports_edit() had the same gap on the read side: it fetched a report by id with no ownership check, so the same user could view any other user's report configuration (recipients, schedule, etc.). 1.2.x already guards both call sites with cacti_authorize_resource(..., 'reports'); develop had lost that check. GHSA-8f5c-3h59-q7gc * security: reject ORDER BY sort_column subquery bypass sanitize_sql_column()'s allowlist regex kept parentheses to support the INET_ATON(col)/NATURAL_SORT_KEY(col) sort wrappers, but that also let MySQL subquery syntax through unchanged, since a subquery like (SELECT(0)FROM(SELECT(SLEEP(5)))a) needs only alphanumerics and parentheses. get_order_string() then detects the '(' and disables its backtick quoting, so the payload reaches ORDER BY verbatim, giving any user with realm 2-5 a blind, time-based SQL injection via sort_column. sanitize_sql_column() now only allows '(' through when the whole value is exactly one of the two known safe wrapper forms around a plain identifier; anything else containing '(' -- including subquery syntax -- now falls back to the default. Added regression tests covering the reported subquery and REGEXP-exfiltration payloads. GHSA-8j73-cvf6-jrvx * security: confine Reports format_file to a bare filename reports_form_save() wrote the POST-supplied format_file value verbatim into reports.format_file, and reports_load_format_file() concatenated it onto CACTI_PATH_FORMATS with no directory-component check, so a value such as ../../../../etc/passwd escaped the formats directory and any Reports-realm user could read arbitrary files readable by the web server. 1.2.x already fixed the same sink (GHSA-g37j-39f4-6r4j / CVE-2026-41884, GHSA-mjvw-mhj5-9jcj / CVE-2026-40084); develop had lost the guard. format_file is now confined to basename() at both the save site and the load site, with the load site additionally rejecting a bare '.'/'..'/empty result before any file IO. Added regression tests. GHSA-5x27-wqpm-m24v * security: authorize Data Source enable/disable against owning device ds_enable()/ds_disable() and the bulk drp_action enable/disable loops in form_actions() passed a user-controlled local_data_id straight into api_data_source_enable()/api_data_source_disable() with no object-level authorization check, so any user holding the Sites/Devices/Data realm could toggle the enabled state of a data source belonging to a device outside their assigned Device Permissions. Added data_source_authorized(\), which allows host-independent data sources (data_local.host_id = 0) and otherwise defers to is_device_allowed(\), and applied it to both the single-item and bulk action paths. GHSA-2fhv-c794-v433 * security: authorize bulk Change Data Source Profile against owning device The bulk 'Change Data Source Profile' action (drp_action == '6') iterated selected_items and updated data_template_data/data_template_rrd/poller_item for each supplied local_data_id with no object-level authorization check, so any Sites/Devices/Data realm user could retarget the RRD step/heartbeat of a data source belonging to a device outside their assigned Device Permissions -- the same class of gap as GHSA-2fhv, in a different bulk action. Guarded with the data_source_authorized() helper added for that fix. This action does not exist on 1.2.x (develop-only feature). GHSA-fc82-w6xq-4jg7 * security: grant AllowOverride All in docker/apache.conf The packaged docker/apache.conf granted only 'AllowOverride FileInfo Options=FollowSymLinks' for the Cacti document root. Cacti's shipped .htaccess deny rules (log/, cache/*, scripts/, cli/, bin/, mibs/, rra/, contrib/, tests/) rely on 'Require all denied' (Apache 2.4, override type AuthConfig) and/or legacy 'Order/Deny' (override type Limit), neither of which FileInfo/Options covers, so Apache silently ignored every one of those deny rules and served the directories over HTTP to any unauthenticated client -- most notably log/cacti.log, which can contain usernames, client IPs, SQL backtraces, and any secrets Cacti logs (SNMP community strings, LDAP bind details). AllowOverride All lets every shipped .htaccess apply as intended. Added a regression test. docker/ does not exist on 1.2.x (develop-only). GHSA-fxfg-j995-j69j * docs: add security advisory / CVE policy to Copilot instructions Documents the policy established for this remediation effort: develop is unreleased, so a develop-only vulnerability does not get a CVE requested; 1.2.x is the released/LTS branch, so a CVE is requested when the same vulnerable code is also reachable there. Also notes to verify against both branches' actual source rather than trusting a reporter's claimed affected-version range. * fix(security): update stale test harnesses and address review findings CI fixes: - CsrfCheckContractProbe.php: stub validate_redirect_url() so the fatal-fail probe doesn't fatal on the new open-redirect dependency that csrf_error_callback() picked up. - InstallerCsrfRecoveryTest.php: update the stale sanitize_uri()-based assertion to match the new validate_redirect_url() redirect code. - psalm-baseline.xml: drop the TaintedHeader suppression for include/csrf.php now that validate_redirect_url() resolves it. Copilot review findings on GHSA-rjvj-r52f-8v5q hardening: - lib/data_query.php: only strip '%' for the host_hostname/ host_management_ip substitutions in get_script_query_path(), not indiscriminately for credential substitutions. - lib/functions.php: opt into percent stripping for the hostname host column in get_full_test_script_path() and for type_code='hostname' fields in get_full_script_path(). - lib/snmpagent.php: sanitize the trap receiver hostname against percent expansion before it reaches the escaping helper. - tests/security/baselines/*.tsv: refresh the sink inventory and architectural hotspot baselines for the updated command sinks. - tests/Unit/DataCollection/Snmp/SnmpCoverageTest.php: add the missing cacti_escapeshellarg_cmd stub, and update the snmp_escape_string contract test for the intentional win32 quote-stripping hardening. - tests/Unit/Security/Shell/CmdMetacharHelperTest.php + tests/fixtures/CmdMetacharHelperWin32Probe.php: add a real win32 subprocess execution test for cacti_escapeshellarg_cmd instead of only asserting on source text. - CHANGELOG: add the pull-request reference to the new entry. * fix(snmp): remove now-unreachable quote-escaping branch in snmp_escape_string The metachar strip on the line above already removes the double quote on win32, so the substr_count(SNMP_ESCAPE_CHARACTER) branch below it could never execute; drop it so lib/snmp.php reaches 100% coverage again. * style: fix pre-existing php-cs-fixer blank-line drift blocking CI data_sources.php and lib/html_reports.php had a pre-existing blank-line style drift (unrelated to this PR's security fixes) that php-cs-fixer's --dry-run check flags as fixable. Since "PHP 8.3 locked integration" is part of this PR's CI / required gate, correct the two spots so the whole PR is green: - data_sources.php: collapse a double blank line before form_actions(). - lib/html_reports.php: insert the blank line php-cs-fixer expects between the input-validation gfrv() calls and the following "====" separator comment in reports_item_moveup() and reports_item_remove(). No behavior change; purely CI/style formatting. * style: refresh phpstan baseline for pre-existing, unrelated findings Three PHPStan findings were masked behind the earlier php-cs-fixer failure and only surfaced once that step passed: - lib/html_reports.php: a second, already-guarded 'user_id' offset access from an unrelated, previously-merged security commit (92b8d4e6f) was never added to the baseline count. - cli/splice_rrd.php, poller_realtime.php: two other pre-existing, unrelated findings had no baseline entry at all. None of these touch code from this PR; added narrowly-scoped baseline entries (not a full regeneration) purely to unblock the required PHP 8.3 locked integration check. * fix(security): address follow-up review findings on consolidated PR Closes the four remaining review findings from the follow-up review round: - package_import.php: trustSigner() now POSTs with a CSRF token instead of $.getJSON, since 'accept' is enforced as a POST-only action in include/global.php's bad_actions list (GHSA-gxgm-8rhv-r7jv). - lib/html_reports.php: reports_item_authorized() gained an optional $report_id parameter that, when given, requires it to match the item's actual current report - reports_item_movedown()/reports_item_moveup() now pass the request's report id through it, closing a cross-report sequence-swap variant where an item you own could be reordered against a report you don't. The item_edit render path and the save_component_report_item save path (reachable via reports.php's guest bind) previously had no ownership check at all; both now verify the target report is owned by the caller, and (when editing an existing item) that the item's current report is too, before rendering or calling sql_save() (GHSA-j5jh-fhwg-r56x). - tree.php + include/global.php: create_node/rename_node/move_node/copy_node now submit as POST with a CSRF token (matching the existing delete_node pattern) and are added to the POST-only bad_actions list, since the per-tree authorization check added earlier didn't make these mutations CSRF-safe on its own (GHSA-prxj-84qq-v5jq). Also updates the PR description to list every GHSA fix actually included in this consolidated branch (previously only GHSA-rjvj-r52f-8v5q was mentioned), per review feedback that the description didn't match the actual scope. * fix: resolve CI failures on PR #8056 (docblock alignment, sink baseline) - lib/html_reports.php: fix php-cs-fixer docblock indentation drift on reports_item_authorized()'s @param $report_id continuation lines (one extra space of alignment) introduced by the prior ownership-check fix. - tests/security/baselines/sink_inventory.baseline.tsv: refresh to track the 4 new header('Location: ' . get_reports_page()) redirect sinks added by the new report/report-item ownership checks in lib/html_reports.php. Verified via tests/security/verify_sink_inventory.sh and confirmed tests/security/verify_architectural_hotspots.sh still reports no new hotspots. * fix: track new PHPStan offset findings from extended reports_item_authorized() phpstan-baseline.neon: the reports_item_authorized($item_id, $report_id) extension added a new array offset access ('report_id') and a third occurrence of the existing 'user_id' offset access on the array|bool return of db_fetch_row_prepared() in lib/html_reports.php. Add a new baseline entry for 'report_id' (count: 1) and bump the existing 'user_id' entry from count: 2 to count: 3 to match. Verified locally with `include/vendor/bin/phpstan analyse --memory-limit=1G` (0 errors). --------- Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> Co-authored-by: Thomas Vincent <thomasvincent@gmail.com> Co-authored-by: TheWitness <copilot@users.noreply.github.com>	2026-09-23 09:45:42 -04:00
Thomas Vincent	d8d81f273d	build: make the containerised test image run the whole suite (#7953)	2026-09-04 13:37:37 -07:00
Thomas Vincent	d4e255b5d9	security: stage nonce CSP migration and gated enforcement (#7662) ... * security: add nonce CSP mode for inline JavaScript protection Reworks the content_security_policy_script setting into a three-mode Inline JavaScript Protection dropdown (HTMLX default, Nonce, None) and ports the per-request nonce, strict-dynamic policy, and CSP violation report endpoint from 1.2.x. The stored option key and its default are unchanged, so existing installs keep their current policy byte-for-byte; nonce mode is opt-in. Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> * security: stage nonce CSP enforcement safely Add report-only migration, gated enforcement, nonce-aware shared emitters, Linux Docker validation, and operator documentation.\n\nRefs #7703 * security: validate CSP header sources Reject malformed alternate-source configurations at the CSP boundary, validate fallback report URIs, and remove the obsolete Psalm header-taint baseline entry.\n\nRefs #7705 * chore: drop the composer.json change from this PR Dependency and manifest changes are a maintainer call and do not belong in this one. Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> --------- Signed-off-by: Thomas Vincent <thomasvincent@gmail.com>	2026-08-16 21:38:46 -04:00
Thomas VincentandTheWitness	5d90e579e2	fix(docker): scope build context to docker/ so it builds standalone (#7513) ... COPY docker/apache.conf and docker/entrypoint.sh only resolve when the build context is the repo root. Building directly from docker/ (docker build .) fails because those paths don't exist relative to that context. Neither COPY step needs anything outside docker/, so narrow the context instead of requiring root. Fixes #7512 Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> Co-authored-by: TheWitness <thewitness@cacti.net>	2026-08-05 10:21:03 -04:00
3 people	242751e41b	fix: harden Docker dev environment (#6621) ... * fix: harden Docker dev environment and add smoke tests Address Copilot review feedback from #6618: input validation on DB env vars in entrypoint, scoped AllowOverride, localhost-only DB port binding, security headers (Referrer-Policy, Permissions-Policy, X-Permitted-Cross-Domain-Policies), reduced php-fpm worker count, expose_php disabled, and .env excluded from git. Adds tests/tools/check_docker.sh — 29-point smoke test covering build, health, PHP extensions, config generation, cron, HTTP, security headers, and DB connectivity. * Update docker/entrypoint.sh Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> * fix: address Copilot and TheWitness review feedback tests/tools/check_docker.sh: - shebang: #!/bin/bash -> #!/usr/bin/env bash (consistent with repo) - add GPL copyright header (2004-2026) - fix grep pattern: single-quoted '^${ext}$' never expanded variable; use double quotes so extension name is actually matched - add opcache to REQUIRED_EXTS (TheWitness) docker/entrypoint.sh: - shebang: #!/bin/bash -> #!/usr/bin/env bash docker/Dockerfile: - rrdtool comment: clarify that 1.9+ enables graph tooltip output via graphv, not merely "hover tooltips" (TheWitness) docker/apache.conf: - X-XSS-Protection: add TODO comment noting it is deprecated and that Cacti sets its own headers at the application layer (TheWitness) --------- Co-authored-by: TheWitness <thewitness@cacti.net> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>	2026-02-17 22:46:45 -05:00
Thomas Vincent	016fbe9b1b	feat: add Docker-based development environment (#6618) ... PHP 8.4-fpm with Apache, MariaDB 11.8, rrdtool. Configurable via .env (ports, credentials, DB tuning). Schema auto-imported from cacti.sql on first startup. Closes #6617	2026-02-14 11:33:49 -05:00