URI validation failure on SVG parsing. Bypass of CVE-2023-23924

Summary

Due to the difference in the attribute parser of Dompdf and php-svg-lib, an attacker can still call arbitrary URLs with arbitrary protocols.

Details

Dompdf parses the href attribute of image tags with the following code:

`src/Image/Cache.php` line 135-150

function ($parser, $name, $attributes) use ($options, $parsed_url, $full_url) {
    if (strtolower($name) === "image") {
        $attributes = array_change_key_case($attributes, CASE_LOWER);
        $url = $attributes["xlink:href"] ?? $attributes["href"];
        if (!empty($url)) {
            $inner_full_url = Helpers::build_url($parsed_url["protocol"], $parsed_url["host"], $parsed_url["path"], $url);
            if ($inner_full_url === $full_url) {
                throw new ImageException("SVG self-reference is not allowed", E_WARNING);
            }
            [$resolved_url, $type, $message] = self::resolve_url($url, $parsed_url["protocol"], $parsed_url["host"], $parsed_url["path"], $options);
            if (!empty($message)) {
                throw new ImageException("This SVG document references a restricted resource. $message", E_WARNING);
            }
        }
    }
},

As you can see from the code snippet above, it respects xlink:href even if href is specified.

$url = $attributes["xlink:href"] ?? $attributes["href"];

However, php-svg-lib, which is later used to parse the svg file, parses the href attribute with the following code:

`src/Svg/Tag/Image.php` line 51-57

if (isset($attributes['xlink:href'])) {
    $this->href = $attributes['xlink:href'];
}

if (isset($attributes['href'])) {
    $this->href = $attributes['href'];
}

Since href is respected if both xlink:href and href is specified, it's possible to bypass the protection on the Dompdf side by providing an empty xlink:href attribute.

Impact

An attacker can exploit the vulnerability to call arbitrary URLs with arbitrary protocols if they provide an SVG file to the Dompdf. In PHP versions before 8.0.0, it leads to arbitrary unserialize, which will lead, at the very least, to arbitrary file deletion and might lead to remote code execution, depending on available classes.