An SSH alternative featuring:
- IPv6 ready
- lightweight, straight forward and extensible protocol using TLS 1.3 or optionally DTLS 1.2 as transport layer
- man-in-the-middle safe due to its authentication mechanism which involves the servers host key into the auth process
- built-in traffic blinding against timing and packet-size info-leak attacks
- not relying on any system auth frameworks such as PAM
- can be entirely run as user, no need to setup config files
- passive/active connects on both ends with most flexible local/remote address+port binding possibilities
- built-in SOCKS5 client side support when doing active connects
- easy to port to embedded systems such as routers
- quiet/hidden mode for secret administration and take-back functionality for owned boxes
- trigger-mode via syslog, mail or other files if requested
- emergency mode to extract all necessary key files from the running binary
- may be started as a CGI with all above functionality, command switches passed via query-string
- integrated tcp-wrapper-like D/DoS protection
- intentionally not passing local $ENV to remote to avoid info leaks
- supports Perfect Forward Secrecy via DH Kex
- can forward TCP and UDP sockets to remote
- SOCKS4 and SOCKS5 support to forward browser sessions to remote
- messenger proxy support
- proxying based on SNI
- SNI hiding mode
- Disguise Filters to mask as different kind of software to global observers
- can use UDP transport mode with DTLS and added reliability and flow-control layer
- transparent roaming support with DTLS client sessions
- suspend/resume support with DTLS client sessions
If you came here for censorship circumvention - once everything is done and working - go to proxywars contrib to learn about how to create WA/TG messenger proxy setups in censorship environments.
Build requires OpenSSL version >= 1.1
or 3.0
or compatible LibreSSL (3.6.1
tested).
Inside the cloned git repo:
$ make -C src
On BSD systems you need to install GNU make and type gmake
instead.
If you have a particular OpenSSL or LibreSSL setup, check the Makefile
and
set the appropriate $SSL
variable. crash builds also nicely with LibreSSL and BoringSSL.
For Android, edit the Makefile.android
or Makefile.android.aarch64
to reflect
your NDK and BoringSSL install and use these. The build was tested with android-ndk-r17b
.
On OSX you want to install OpenSSL via brew install openssl@1.1
or by hand before make
.
On Windows you need to install cygwin and select
the appropriate gcc, gcc-g++, perl, openssl (1.1.1), libssl (1.1.1), libssl-devel (1.1.1), make
and git
packages before the build in order to clone and make
this repo. Make sure
your openssl versions for the tool itself, the runtime libs and devel package are all
the same.
crash was successfully tested on Linux, FreeBSD, NetBSD, OpenSolaris, OSX and Android.
After that, to generate the required server and authentication keys:
$ make -C src keys
or see further instructions in this document. If you want to use ephemeral keying (aka PFS), invoke
$ cd src; ./newdh
before make
in order to generate DH parameters before the build. Thats not strictly necessary
as of TLS 1.3, since the Kex will most likely chose one of the ECDH variants, but if you customize
your setup, it is recommended to generate your own DH params.
If make
detects that TLSv1.3 is not available on the system or TLS_COMPAT_DOWNGRADE
is
defined, the binaries are built with TLSv1.2 only. This is to allow using it on legacy
systems when no other options are available. Obviosuly, the built binaries are not
compatible to normal builds, but include full support of all other features.
The OpenSSL 3 API is quite different from the OpenSSL-1.1 API. In order to make
use of OpenSSL 3, you have to edit Makefile
and newdh
to reflect your path setup
for your OpenSSL install. Invoking newdh
is mandatory, unlike for the 1.1 builds. After
that you just do make
and everything should be the same as with 1.1.
proudly sponsored by:
crash does not need any config or option files to run. Its easy and straight forward to use. Anything can be enabled/disabled by runtime switches:
stealth@linux ~> ./crashd -h
crypted admin shell (C) 2024 Sebastian Krahmer https://github.com/stealth/crash
Usage: ./crashd [-U] [-q] [-a] [-6] [-D] [-H host] [-p port] [-A auth keys]
[-k server key-file] [-c server X509 cert] [-L [ip]:port] [-S SNI]
[-t trigger-file] [-m trigger message] [-e] [-g good IPs] [-N] [-R]
[-x socks5://[ip]:port] [-G method:prefix:action] [-w]
-a -- always login if authenticated, despite false/nologin shells
-U -- run as user (e.g. turn off setuid() calls) if invoked as such
-e -- extract key and certfile from the binary itself (no -k/-c needed)
-q -- quiet mode, turns off logging and utmp entries
-6 -- use IPv6 rather than IPv4
-w -- setproctitle to `[kthreadd]` (must be last arg!)
-H -- host to connect to; if omitted: passive connect (default)
-p -- port to connect to when active connect; default is 2222
-L -- local [ip]:port used for binding ([0.0.0.0]:2222)
-g -- file containing list of good IP/IP6's in D/DoS case (default off)
-A -- authorized-key file for users if starts with '/'; folder inside ~
containing authorized_keys file otherwise; 'self' means to use
blob-extraction (see -e); default is .crash
-k -- servers key file; default is ./serverkey.priv
-c -- X509 certificate-file that belongs to serverkey (-k);
default is ./serverkey.pub
-t -- watch triggerfile for certain message (-m) before connect/listen
-m -- wait with connect/listen until message in file (-t) is seen
-N -- disable TCP/UDP port forwarding
-D -- use DTLS transport (requires -S)
-x -- use this SOCKS5 proxy when using active connect
-R -- allow clients to roam sessions
-G -- Traffic Disguise Filters, check docu
-S -- SNI to hide behind
Most of it is pretty self-explaining. crashd can run as user. -U
lets crashd
skip setuid()
calls, effectively being able to run as user. In this case, it only accepts
logins to that user then by checking login name's uid
against current euid
.
Both, crashc and crashd can use active and passive connects. Whenever
a host-argument -H
is given, it uses active connect to this host
and the belonging port -p
. It also accepts -L
which specifies the local address and port it has
to bind to, either before doing active connect (-H
) or passively (no -H
given).
This way - from TCP point view - client and server role may be reversed, while still having
crashd as the shell server.
If -w
is used it forks itself as [kthreadd] and tries to wrap around its
pid
to be somewhere around the system daemons. As -w
is overwriting main()'s argv
array,
it must appear last in the option list, otherwise option processing will not work
correctly. You can set the process name as TITLE
def inside the Makefile
.
For testing, when you did make keys
(next section), you can just run
src $ ./crashd -U -L [127.0.0.1]:2222
(or omit -L
paramater to bind to the default port on any address) and
src $ ./crashc -v -K none -i authkey.priv -H 127.0.0.1 -p 2222 -l $USER
Unless you want to use SNI-hiding (see section below), you can type straight ahead:
$ make -C src keys
But you can also do it by hand. To generate a X509 certificate containing the server key:
$ umask 066
$ openssl genrsa -out serverkey.priv 4096
$ openssl req -new -x509 -nodes -sha1 -key serverkey.priv -out serverkey.pub
To extract the public key in a form crashc can use it as a hostkey for comparison:
$ openssl x509 -in serverkey.pub -pubkey -noout > HK_127.0.0.1
So you have HK_127.0.0.1
as the known-hosts keyfile for this host.
As an alternative, you can use crashc with -v
upon connect to
extract the pubkey. But note that this might already be a key presented to
you during an attack. So only do that if you know that the connection is
not tampered with (e.g. single user on localhost).
Unless you use SNI hiding (see section below), the values you enter for Country-Name, Email, CN etc. do not matter since crashc is not validating the X509. It just compares the public key value it obtained from the server with the key it has in its local key-store belonging to that server (similar to SSH). The server key is not encrypted since crashd is usually started via init scripts. Instead, the key file must have proper permissions so only appropriate users can read it (mode 0600). You can, if you like, also encrypt the server key but then you have to enter a pass-phrase whenever crashd is started.
To generate a public/private RSA keypair for your authentication:
$ openssl genrsa -out authkey.priv -aes256 4096
$ openssl rsa -in authkey.priv -pubout -out authkey.pub
Copy authkey.pub
to ~/.crash/authorized_keys
on the remote box, and
use authkey.priv
for the crashc -i
argument. Note, that upon authentication
you will be asked for the pass-phrase to unlock your private key that is
stored locally. The pass-phrase will not travel the network.
Auth-Key sizes larger than 7500 bit must not be used; the tokens do not fit into the auth handshake otherwise.
crashc is using the .crash/
subdir by default to check for
already seen server keys. If you connect to a host via -H $host -p $port
,
a keyfile of form .crash/HK_$host:$port
is looked up unless you specify an
absolute path to a known keyfile.
By default, crashc will compare server hostkeys to the local key cache
that is found inside the .crash/
subdir of CWD. You may override the path of the cache
folder by using the -K
switch. For example by using -K ~/.crash/
, you use the folder
inside your home directory. If the pathname does not end with a slash, it is treated as
a filename instead of a directory. If a cache directory is used instead of an
filename, each hostkey is expected to be found inside the folder as of the name HK_$HOST:$PORT
where $HOST
is the -H
argument and $PORT
the -p
argument. If using -v
the server hostkey will be printed on stderr
and may be pasted to the cache folder
into the HK_$HOST:$PORT
file.
Hostkey checking may be suppressed by using -K none
.
The crash auth protocol incorporates the server host key when signing authentication
requests. This way its not strictly necessary to check server host keys as
you know it from SSH password authentication. Two things have to be considered
if host-key checks are suppressed with -K none
though:
- The user-name will potentially leak to a MiM server
This is not an issue if you use a system user-name such as root
.
- The MiM could sort of phish you, by showing you a fake-shell where
you think it belongs to your real server. This could be used to
wait for
su
and similar commands and to record sensitive information as you type on the MiM shell.
To conquer this, you have to make sure you are indeed on your real shell
when you see the prompt. This can be achieved by echoing a secret token
to the tty upon login, for example via one of the .profile
or .bashrc
files. As the MiM cannot know this token, you can be sure you have a
confidential and untampered session when you see this token upon login;
even if you omit the host-key check.
As crashc is not transfering env vars to the remote side for a reason, keep in mind
that certain stuff is unset, such as $TERM
. I.e. if you want to run an editor on the remote
shell and started crashc from within an xterm, you have to TERM=xterm vi file
in order
to have a useful editing session. Likewise for other programs that you expect to work and
require specific environment setup.
crashd automatically detects whether it has been invoked as a CGI by
a web-server by checking QUERY_STRING
environment variable. It parses
and converts the query-string into arguments it understands. It
does not translate %2F
etc characters! They should not be needed,
since spaces, '(' and other weird characters do not make sense when
calling crashd. Arguments that don't have a parameter such as -U
have to be given =1
argument to enable it, such as in:
http://127.0.0.1/cgi-bin/crashd?-K=/path/to/serverkey.pem& \
-C=/path/to/pubkey.x509&-p=1234&-A=/tmp/.crash/authorized_keys&-U=1&-a=1
which invokes crashd on the host 127.0.0.1 as user (probably "wwwrun" or whatever the web-server is running as).
For pen-testing or for emergency case, crashd has the -e
option.
If -e
is used, it extracts the server key-file and the X509 certificate
from the ELF binary file, which have to be appended before using -e
:
$ cat serverkey.priv>>crashd
$ cat serverkey.pub>>crashd
$ cat authkey.pub>>crashd
The order of appended keys is important.
If you give -A self
instead of a valid authentication directory or file,
crashd also extracts the user-key used for authentication from its binary.
The keys are extracted from the binary at runtime and stored in temp files
of pattern /tmp/sshXXXXXX
(/data/local/tmp/sshXXXXXX
on Android).
Make sure to erase them securely upon last login, since they contain private keying
material.
This is useful in pen-tests where you cannot upload arbitrary amount of files or you do not know the exact pathname of the upload storage:
$ curl 'http://127.0.0.1/cgi-bin/crashd?-A=self&-U=1&-e=1&-a=1'
-a
is needed since most likely the wwwrun user has a /bin/false
shell,
which -a
ignores.
crashd is using mkstemp()
to store the key files temporarily (just see above), with
mode 0444 (world readable) since it needs to access authentication
files as user. So be warned that, if you have users, they may read
the private key used during SSL handshake. After all, its just an
emergency mode. Stripping the crashd binary is not possible after
appending the keys, or they will get lost.
Back-connect etc. also work in CGI mode as well.
If using that, client should use -K
switch to tell client which key to use
to authenticate the server.
crash uses the same network engine as psc. Therefore
you may use the same -U
and -T
parameters as known from psc and which are similar
to those of OpenSSH's -L
parameter. It will bind to lport
and will forward connections
to [ip]:rport
, initiating the connection from the remote host. The same works for UDP
packets, which is not possible with SSH.
If you are interested in messenger proxy setups in copland countries, you can check contrib
folder.
crash also supports forwarding of TCP connections via SOCKS4 (-4 port
) and SOCKS5
(-5 port
). This sets up port as SOCKS port for TCP connections, so for instance you
can browse remote networks via crashc sessions without the need to open any other
connection during a pentest. If you pass -N
to crashc, it enables DNS name resolution
on the remote side, so you can also use chrome with it. But be warned: There is a privacy
problem with browsers that try to resolve a sequence of DNS names upon startup that
is not under your control. Also, if your remote side has a broken DNS setup, your typing
shell may block for several seconds if DNS reply packets are missing. There are no good
async resolver functions which are embeddable and portable so I had to rely on
getaddrinfo()
in the single thread at the price of possible blockings for several seconds
if DNS problems exist. Thats why name resolving has to be enabled explicitly. crashd
tries to minimize this potential problem with DNS lookup caches though, so in most
situation it should just work painlessly.
If you pass -X IP-address
(must come before any other proxy argument), you can bind your local proxy
to an address different from 127.0.0.1
, so you can share the proxy in your local network.
There is also a client side SOCKS5 support available when using crashc with -x
.
In some circumstances you might want to change the endpoint of the proxy session based
on a SNI that you receive in the TLS ClientHello
. For convenience, crashc integrates
support for that by using -Y lport:SNI:[ip]:rport
which listens on lport
and forwards the
given SNI
to ip:rport
. A fallback of default
SNI can be given so that any non-matches or
missing SNIs will be forwarded to that destination.
crash allows to use all of its trickery above also on a DTLS 1.2 transport layer based
on UDP. I have added basic flow control and reliability, so you can even xfer files and use
port forwarding as with TLS 1.3. The reason for adding DTLS is that some countries have
TCP egress filters that only allow incoming connections. It is harder for censors to tell
which UDP packets establish an outgoing connection, as there is nothing like a "connection"
with UDP. With DTLS sessions, which are established by the -D
switch on both sides, a SNI
is mandatory.
When forwarding UDP ports on DTLS sessions, make sure you will not send UDP payloads larger than
1320 bytes across the sockets, as it is necessary in UDP case to keep enough room for headers
and record layer without the need to fragment the packet, as DTLS honors packet boundaries
(there is nothing like a stream as in TCP, just datagrams).
DTLS mode is still experimental (although working stable) and will switch to DTLS 1.3 as soon
as it is implemented widely (DTLS 1.3 RFC was just finished 2022).
This is an experimental feature, although working stable.
When using DTLS sessions and crashd is started with -R
, you will get the following:
- transparent roaming of the client sessions - including existing SOCKS connections - which allows to switch underlying physical layer, VPN, Interface, NAT or IP address without even noticing it
- crashc may be terminated via
SIGTERM
, so it will dump the session to a ticket file (-t
) which can later be resumed from by passing the correct dst IP:port and ticket but w/o the need to authenticate again (no-i
) - with full roaming support
In the 2nd case, the ticket file will not be encrypted, so make sure you never leak it. This allows you to switch off your laptop and continue working from elsewhere or even share the ticket to another admin who then continues your session.
One thing is special with regards to bound server ports when using roaming: Due to
UDP internals, the next open session for a followup "connect" will be on the next
free port in the range of [port, port + 1000]
and not on the same port as when using TCP.
This needs to be as with roaming we cannot actually call connect()
to virtually create a
connected tuple, as the next session packet can arrive from anywhere - not just from the
originating IP as happens with TCP. So when you start the server with -p 2222
and one
roaming session already exists, the next one needs to "connect" to port 2223
. If the
session at port 2222
is finished (not suspended, but really finished), port 2222
will become available again to the next client.
Suspend/Resume does not work yet with LibreSSL builds, but roaming does.
Traffic analysis mitigation has differant goals:
- make it hard to find out actual typing sequences and potential info leaks about whats being typed
- make it hard for a global observer to track connection streams across packet mixes or hubs
- make it hard for censors to identify/distinguish crash sessions from a set of "legit" connections
It is not possible to reach all three goals at the same time, e.g. you want randomized packet sizes to make it hard for observers to know you are using a crash session, but this will also make your packet stream unique across mixes.
Completely mitigating traffic analysis for a capable (global) observer is very hard.
It would require many crash users so to sink all individual packets in a swarm and
make it impossible to find patterns that could be used to track individual users across
packet mixes. It would also require a fixed packet size for all packets as well as a
constant delay between the sends to make all connections look equal. Even then, there's
still the problem of the overall amount of traffic sent that may be measured and used
to track individuals. As having constant size and delays would make the connection
feel slow or even unusable, crash lets you choose between traffic policies which are
controlled by -R <level:factor>
at client side. Level is an integer with the following meaning:
-
0: disable all padding of payloads and do not inject random traffic
-
1: pad payload to rand size up to 1320 byte boundary, no injects
-
2: pad payload to rand size up to 1320 byte boundary, random injects client side
-
3: pad payload to rand size up to 1320 byte boundary, random injects with server responses
-
4: pad payload to the next 256, 512, 1024 or 1320 byte boundary, no injects (default)
-
5: pad payload to the next 256, 512, 1024 or 1320 byte boundary, random injects client side
-
6: pad payload to the next 256, 512, 1024 or 1320 byte boundary, random injects with server responses
-
7: pad payload to 1320 byte boundary, no injects
-
8: pad payload to 1320 byte boundary, random injects client side
-
9: pad payload to 1320 byte boundary, random injects with server responses
Factor is a multiply factor 1..100
that adds as many NOP packets per real packet in order
to make it harder to match amount of input traffic to output traffic on proxy hosts.
1320 is crashds internally used MSS. The values were chosen in a way so that sent data fits most likely into a single packet. Note however that these are the packet sizes (plus the TLS record size) as it is passed to the TCP stack. TCP will decide itself how it will send the segments. There is no way to enforce 'TCP packet sizes', but this does not matter as the deps to the actual payload size is already blurred.
A higher Level does not automatically mean a better analysis mitigation. You have to chose the best
-R <level:factor>
depending on your personal needs.
In DTLS mode there are always ping packets in order to implement synchronization and flow control.
If you live in a country with restrictive egress filtering, it may be helpful to test how long
connections can survive. Note that due to -4
and -U
which allows to proxy TCP and UDP (DNS)
to a remote site, crash may be used as a shadowsocks alternative that requires
basically no setup and just needs a user-shell behind egress.
If you think that all of this is paranoia, go get some product sheets for devices that detect and classify SSH traffic by behavioral analysis.
By default, the crashd will show a banner upon connect to tell the peer major and minor version
numbers. Censorship countries might block addresses which show banners they dislike. To combat this,
crash allows for a TLS-only mode that is indistinguishable from a HTTPS session. Just start
crashd with -S
and give a semi-secret name (Server Name Indicator, SNI). Only clients that also
use the correct -S
parameter will reach the gate for authentication at all. Other TLS sessions
will just be rejected. Note that the SNI travels the network in plain-text and that -S
is not meant
for authentication. The only reason for SNI hiding is to hide the crash banner from probing/crawling.
You may also use SNI proxies such as sshttp to hide crash even
deeper and to forward all non-correct SNI connects to some web-site. This way you may hide your server
behind neutral web-sites from aggressively probing/blocking censors.
In order for probing to not reveal that you are running crash by checking the X509 certificate
details, you should use reasonable values for Country Name, City etc. when asked for it during
the make keys
process. For instance it would make no sense to setup a pro-regime web-site
to hide behind and enter anti-regime values for the X509 specific naming.
Inside the contrib
folder you will find a nginx config file that you can integrate into
your setup along with comments how you would create a connect from outside to your nginx server
in order to have a crash session based on a SNI that you chose.
Taking the feature of SNI hiding one step further. Some countries use network data gathered at their border routers to scan destination machines and check whether the content or software there could pose a threat to their leaders. It is therefore not good to always tell anyone openly that a crashd is running on a certain port, even if the peer shows up with the right SNI, as the SNI could have been sniffed by a global observer. Entering Disguise Filters.
Upon connect, you have to show up with a correct (pre-)secret in order to start a crash session. This can't be known by an observer as its hidden inside the TLS stream (unlike the SNI). If the secret is not correct, crashd disguises as another - innocent looking - software.
Currently, there is only one Disguise Filter, redirect1
, which masks as a web server
sending a redirect of your choice. Disguise Filters always also require -S
:
$ ./crashd -L [0.0.0.0]:4433 -c serverkey.pub -k serverkey.priv \
-G redirect1:mydirtysecret:https://www.ccc.de -S localhost
So only those who know can start a shell session:
$ ./crashc -H 127.0.0.1 -p 4433 -l stealth -i authkey.priv -S localhost -G mydirtysecret
All others, e.g. curl https://localhost:4433 -k -v -L
will be redirected to https://www.ccc.de
.
(where localhost
was just chosen for testing to make curl have the right SNI)
When a Disguise Filter is triggered, you will see it in the logs. This also allows admins to
have shell servers reachable from outside which just map to the legit web server when not
prompted with the correct (pre-) secret.
For sure; for a disguise to work against censors with a large dick, your story has to be
perfect, i.e. the CNs etc. of the certificate have to look legit, even better signed by
a legit CA and the redirect has to look reasonable.
Although there is nothing like sftp
for crash, it may be used for file up/downloads.
In order to upload a file:
~ $ crashc -H host -i authkey.priv -l root -c 'dd of=/path/on/remote status=progress' < local.file
Or to download a file:
~ $ crashc -H host -i authkey.priv -l root -c 'dd if=/path/on/remote status=progress' > local.file
Note that in the download case you must not specify the -v
switch since this would add
the verbose output to the local.file
. For -c
commands, crash will forward stdout
and
stderr
separated to the local tty's fd 1 and 2, so above commands add a nice progress bar
during the xfer.
crash is assuming a MTU of 1500 and using a MSS of 1320, so that TLS record layer and some other
meta data fits into this MTU and even into network devices with smaller MTU ~1400, which should fit with most
VPN setups etc.. If you are using DTLS mode (that is UDP) and a VPN or whatever with a much smaller MTU,
you might want to compile crash with lower values which you can change in misc.h
. If you are using TCP,
i.e. not using the -D
switch, these values do not matter for you.
crashd includes some sort of D/DoS protection. Only one connection per second is allowed per IP, except if the IP is listed (or the network it belongs to) in a good-IP file given with -g at startup. Per default no good IPs are assigned. Network-address-goodness only works with IPv4 yet. A simple good-IP file may look like this:
# sample good-IP file
192.168.3.1
192.168.2.0
10.0.0.0
fe80:216::1234
# end of file
Together with the interval timer for hanging un-authenticated connections this allows to have no more than 12 'hanging' crashd's at the same time, still allowing you to login if you are listed in good-IPs and your underlying TCP/IP stack is not already trashed.