mirror of
https://github.com/Cacti/cacti.git
synced 2026-10-23 03:27:02 +09:00
1e5329dd89
* 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, $report_id) extension added a new array offset access ('report_id') and a third occurrence of the existing 'user_id' offset access on the array|bool return of db_fetch_row_prepared() in lib/html_reports.php. Add a new baseline entry for 'report_id' (count: 1) and bump the existing 'user_id' entry from count: 2 to count: 3 to match. Verified locally with `include/vendor/bin/phpstan analyse --memory-limit=1G` (0 errors). --------- Signed-off-by: Thomas Vincent <thomasvincent@gmail.com> Co-authored-by: Thomas Vincent <thomasvincent@gmail.com> Co-authored-by: TheWitness <copilot@users.noreply.github.com>
30 lines
1.0 KiB
XML
30 lines
1.0 KiB
XML
<?xml version="1.0" encoding="UTF-8"?>
|
|
<files psalm-version="6.16.1@f1f5de594dc76faf8784e02d3dc4716c91c6f6ac">
|
|
<file src="include/global_session.php">
|
|
<TaintedHtml>
|
|
<code><![CDATA[gfrv('action', FILTER_VALIDATE_REGEXP, ['options' => ['regexp' => '/^([-a-zA-Z0-9_\s]+)$/']])]]></code>
|
|
</TaintedHtml>
|
|
<TaintedTextWithQuotes>
|
|
<code><![CDATA[gfrv('action', FILTER_VALIDATE_REGEXP, ['options' => ['regexp' => '/^([-a-zA-Z0-9_\s]+)$/']])]]></code>
|
|
</TaintedTextWithQuotes>
|
|
</file>
|
|
<file src="install/functions.php">
|
|
<TaintedFile>
|
|
<code><![CDATA[$cacheFile]]></code>
|
|
</TaintedFile>
|
|
<TaintedHtml>
|
|
<code><![CDATA[$data]]></code>
|
|
</TaintedHtml>
|
|
</file>
|
|
<file src="lib/functions.php">
|
|
<TaintedHtml>
|
|
<code><![CDATA[$omessage]]></code>
|
|
<code><![CDATA['Message: ' . trim($string) . PHP_EOL]]></code>
|
|
</TaintedHtml>
|
|
<TaintedTextWithQuotes>
|
|
<code><![CDATA[$omessage]]></code>
|
|
<code><![CDATA['Message: ' . trim($string) . PHP_EOL]]></code>
|
|
</TaintedTextWithQuotes>
|
|
</file>
|
|
</files>
|