Implement a command-only session type over the HrrRbSsh client
Connection Channels' file descriptors, adjust from base command
session to deal with the separate reader/writer IOs. Technically,
a TTY session works out of the box here as well.
Implement a pair of showcase Ruby payloads using net/ssh to call
back to the handler, create a shell channel, and loop piping I/O
between framework session and client via the Ruby backtick exec.
Next Steps:
Command payloads need to be written for every major interpreted
language as well as some sort of bashism a la openssl_double if
it comes to that, but preferably single socket implementation.
Testing:
Very minimal, needs a good run through by the community and R7
Create the Rex::Proto::Ssh namespace and implement hrr_rb_ssh
objects in the Rex namespace with Rex' sockets stack, permissive
default connection options to accept any authentication, ACLs for
port forwarding, and a modified Connection #initialize method for
simplified instantiation.
The actual Rex::Proto::Ssh::Server object follow standard Rex
semantics for services to permit use in handlers, exploits, and
auxiliary modules in the same manner as the Http::Server is used
today.
This work is far from complete - the HrrRbSsh low-level objects
should be decomposed into Rex' pattern as they currently depend
heavily on their own Procs and Threads internally which use the
parent Connection object's attributes to look up which Proc should
be called in response to a Channel or Authentication request. This
is a bit difficult to do piecemeal given the intertwined data and
execution dependency inside of HrrRbSsh.
Next steps:
1. Create handler and reverse_ssh command session payloads
1a. POSIX systems generally have SSH already
1b. Windows Powershell is likely able to hotload SSH libs as a
pre-stager.
1c. Other interpreted runtimes usually have client-ssh libraries
2. Implement options parsing, handling, etc for user-customizable
functionality from the Msf namespace.
3. Use libssh2 to permit mettle use of the SSH transport.
3a. Bother @OJ about doing the same for Windows Meterpreter.
3b. Bother @zeroSteiner to implement SSH transport for PyMeterp.
3c. Ask @timwr and @mihi how viable this is for Android/Java.
3d. See if @OJ will do this for CLR-meterp on a livestream.
4. Write a post module which runs this SSH server from a mettle
session using the compromised hosts's own host keys and passwd
file for "valid" authentication while logging and MITMing the
entire command stream remotely - help folks shake off some of the
complacency around "secure shells."
5. Write a plugin which will permit sharing of the entire console
context, specific acquired sessions, or TCP forwards with the
pugin's SSH clients based on the credentials used to connect.
6. Further decompose the server code into Rex to permit writing
client fuzzers, loggers, and other tooling to test the posture of
SSH client implementations.
7. Work to implement SSH client functionality in HrrRbSsh and
convert the rest of Rex/Msf to use this library instead only.
8. Rewrite and import MetaSsh into Metasploit proper using the
new Rex::Proto::Ssh code.
-----
Huge thanks to @hirura for writing HrrRbSsh. I've spent countless
hours on and off working to implement server-side semantics in
net/ssh years ago and it was becoming a pretty full rewrite due
to the ordering of server/client req/resp messages and who sent
what to whom. This library is a much much cleaner implementation
and provides full, modern SSH servers in pure Ruby.
This gives us the ability to force TLV encryption if for some reason
it's not already in place, and it means we can renegotiate a new key on
the fly if we want to.
According to https://tools.ietf.org/html/rfc7230#section-3.3.2, a zero content-length is valid for some kinds of HTTP methods.
Instead of implicitly disabling auto_cl if there is no actual content, disable auto_cl default for HTTP methods where semantics of the message do not anticipate any content. This can still be overridden by a caller if it still wants to add an empty content-length for HTTP methods where it does not normally make sense (e.g. if it exploits a bug.)
expand_path is not implemented consistently across platforms and
sessions, which leads to confusing behavior. In places where we have trivial
single variable expansions, this changes modules and library code to just use
getenv.
We'll look at the rest individually to see if they can also be reimplemented in
terms of getenv.
William Vu