Guardian Agent: secure agent forwarding for Mosh or SSH
Guardian Agent (now in beta) allows users to securely empower remote hosts to take actions on their behalf, using their SSH credentials. It allows Mosh and SSH users to enable agent forwarding for every connection, even to hosts they may not fully trust.
Guardian Agent is an alternative to traditional
forwarding, which can only safely be enabled when connecting to
trusted hosts. The traditional
ssh-agent protocol doesn't give
the agent information about which host is asking to perform a command
on the user's behalf, which server that hosts wants to connect to,
or which command the host wants to perform:
By contrast, Guardian Agent provides more-constrained agent forwarding that can safely be enabled on any connection. It can be used alongside Mosh or SSH.
To use Guardian Agent, a user runs
sga-guard on her local machine to establish
a secure forwarding channel to an intermediary host (e.g. in EC2). On
the intermediary machine, she can then use
sga-ssh as a drop-in
ssh. The local
sga-guard verifies the identity of
(a) the intermediary host, (b) the remote server, and (c) the
command*, either by prompting
the user or based on a stored policy. No changes to the server are
required, but full functionality is only available when connecting to
servers running OpenSSH.
Based on feedback to this beta/prototype, we may incorporate the techniques behind Guardian Agent more deeply into Mosh.
How does Guardian Agent work?
Guardian Agent allows an SSH client (
sga-ssh), running on a
partially trusted machine, to request the SSH agent (
running on a trusted machine, to execute commands on an SSH server
(e.g., GitLab), such that the identity of the server as well as the
SSH session command can be verified by the SSH agent, with the
server's own code unaltered.
The scheme works by first having
sga-ssh (on the intermediary host)
request that the user's agent allow it to run a particular command on
a particular server. The user approves or denies the request, or the request
is auto-approved according to a pre-existing policy. (These policies are stored
sga-ssh then establishes a TCP connection to the
server, and securely tunnels it back to
establishes an end-to-end authenticated SSH connection directly with
the server, over the forwarded TCP connection. The local machine
allows the delegate to issue the approved command, and then to
``lock'' the session by using the
no-more-sessions request, so that
no more commands may be issued on this connection.
After the session is locked,
sga-ssh to rekey the
session, taking over in the middle of the connection, and reusing the
existing TCP connection to the server. From this point,
sga-ssh can complete the operation, with data flowing directly between the intermediary and the server (as in traditional
Warning: this tool is in beta and is intended as a technology prototype. It was first released in October 2017. Feedback is greatly appreciated, but please use at your own risk.
Using Guardian Agent requires installation both on your local machine (the
one with your SSH private keys) and on each of the intermediary machines you
want to securely forward
ssh-agent to (the machines on which you want to run an
SSH client without having the keys on them). No installation is required on the
sudo apt-get install openssh-client autossh ssh-askpass curl -L https://api.github.com/repos/StanfordSNR/guardian-agent/releases/latest | grep browser_download_url | grep 'linux' | cut -d'"' -f 4 | xargs curl -Ls | tar xzv sudo cp sga_linux_amd64/* /usr/local/bin
CentOS / Fedora installation
sudo yum install -y mosh openssh-clients autossh openssh-askpass curl -L https://api.github.com/repos/StanfordSNR/guardian-agent/releases/latest | grep browser_download_url | grep 'linux' | cut -d'"' -f 4 | xargs curl -Ls | tar xzv sudo cp sga_linux_amd64/* /usr/local/bin
brew tap theseal/ssh-askpass brew tap theseal/guardian-agent brew install guardian-agent
Make sure Guardian Agent is installed on both your local and intermediary machine.
On your local machine
Start guarded agent forwarding to the intermediary machine:
[local]$ sga-guard <intermediary>
You should then expect to see the following message:
[local]$ sga-guard aws-ubu Connecting to aws-ubu to set up forwarding... Forwarding to aws-ubu setup successfully. Waiting for incoming requests...
Guarded agent forwarding is now enabled on the intermediary.
On the intermediary
Connect to the intermediary (e.g., using standard ssh or mosh).
To enable several common tools (scp, git, rsync, mosh) to use the guardian agent instead of the default
[intermediary]$ source sga-env.sh
You can also add this line to your
~.zshrc/... file on the intermediary hosts.
You can then use
mosh as you would normally do.
[intermediary]$ git clone email@example.com:user/repo ... [intermediary]$ scp foo.txt remote-host:Documents/foo.txt ...
You can also use
sga-ssh as a drop-in replacement to an ssh client:
[intermediary]$ sga-ssh <server> [command]
Command verification requires the server to support the
extension. This extension is present on OpenSSH servers, but
unfortunately not implemented on other SSH servers (including GitHub, which uses
executing a command on a server that does not support this extension, only the
identity of the intermediary and the identity of the server can be constrained and verified by the agent
(but not the contents of the command).
Guardian Agent supports two types of interactive prompts: graphical and
terminal-based. The graphical prompt requires the
DISPLAY environment variable
to be set to the appropriate X11 server.
If running in a terminal-only session (in which the
variable is not set), a textual prompt will be used instead.
Customizing the SSH command
sga-guard, the default SSH client on the local machine is used to
set up the connection. This requires
ssh to be found in the user's
specify an alternative SSH client or specifying additional argument to the
client, use the
--ssh command-line flag.
sga-stub is not installed in the user's
PATH on the intermediary
machine, its location must be specified when setting up secure agent forwarding
from the local machine:
[local]$ sga-guard --stub=<PATH-TO-STUB> <intermediary>
Building from Source
- Install go 1.8+
- Get and build the sources:
go get github.com/StanfordSNR/guardian-agent/...
- Copy the built binaries (
$GOPATH/binto a directory in the user's PATH.
- Copy the scripts
$GOPATH/src/github.com/StanfordSNR/guardian-agent/scripts/sga-env.shto a directory in the user's PATH.
In case of unexpected behavior, please consider opening an issue in our issue tracker.
We'd also greatly appreciate if you could run the tool in debug mode by setting the
--logfile=<LOG-FILE> flags and attach the log file to the issue.
ssh-agent forwarding really insecure? What is the point of Guardian Agent?
A: The ssh(1) man page warns that "Agent forwarding should be enabled with caution," because
ssh-agent protocol does not allow the agent to verify (a) which intermediary machine is making the request, (b)
which remote server the intermediary wants to authenticate to, or (c) what command the
intermediary plans to run on the remote server. The agent signs a blank check---
an opaque challenge from an unknown server that will allow the intermediary to execute
any sequence of commands on the user's behalf.
Guardian Agent is a prototype of a system for secure agent forwarding that could safely be enabled on every outgoing connection, because the local agent can verify and enforce security policies regarding who wants to do what to whom.
Q: What's the exact attack scenario where I could get burned with traditional
A: A compromised or malevolent intermediary can send rogue challenges and use the user's identity to authenticate to other servers or to run unauthorized commands. A user that enables ssh-agent forwarding will be asked yes or no on signing "something," and may think she is allowing an EC2 machine to run "git push" to GitHub. Unbeknownst to the user, she is allowing a different machine (that she is also logged in to) to connect to some other sensitive server that she has permissions on and do something evil.
Q: What if I only use
ssh-agent forwarding when I SSH to intermediaries that I trust?
A: If the user trusts the software and system administrators on the intermediary host, it is probably fine to use ssh-agent forwarding as it exists today. However, with this level of trust, it may also be fine to simply place a private key on the intermediary's hard drive. Guardian Agent is a prototype of an approach that could safely be enabled by default.
Q: Can I use Guardian Agent to constrain an intermediary to only pull from (or only push to) a limited set of remote Git repositories?
A: Yes, if the remote Git server is running an SSH server (such as OpenSSH) that supports the
no-more-sessions extension. (The name of the repository, and the difference between pulling and pushing,
are both represented in the command.) Among popular Git-hosting services that we
are aware of, currently only GitLab
appears to support this currently. GitHub and Bitbucket use other SSH implementations and
do not allow Guardian Agent to constrain the intermediary to only push or pull from certain
Q: Is Guardian Agent secure?
A: Guardian Agent is a beta that was first released for beta testing in October 2017. It has not accumulated enough testing and scrutiny to make claims that the implementation is bulletproof.
Q: What language is Guardian Agent written in?
A: Go! Guardian Agent is a beta intended to solicit feedback from the community, especially with regards to the basic design of a secure protocol that works with unmodified remote SSH servers. We found Go and the Go SSH library to be helpful in prototyping this tool.
Q: What is the connection to Mosh (mobile shell)?
A: Many Mosh users have asked for ssh-agent forwarding support. Guardian Agent was developed by some of the Mosh developers and can be used with Mosh today. Based on feedback to this prototype, we may integrate Guardian Agent more fully into Mosh as a system for secure ssh-agent forwarding that is safe enough to leave on by default.
Q: Who wrote Guardian Agent?
A: Guardian Agent was developed by students and faculty in the Stanford University Department of Computer Science (Dima Kogan and Henri Stern, advised by Keith Winstein and David Mazières).
Q: Where should I send feedback?
A: Please file an issue on GitHub.
Q: Where can I read more about how Guardian Agent works?