When you're forwarding an agent the material itself is kept on your local client, but if someone has root on the server that your agent is forwarded to, they ca

When you're forwarding an agent the material itself is kept on your local client, but if someone has root on the server that your agent is forwarded to, they can hijack your ssh-agent socket and ssh to any other server "as you", and your local ssh-agent will comply with the request and forward the key. There's a "lesser of two evils" argument you could make: by using a separate key that only has rights to git, you're protecting your "main" ssh identity from being hijacked. So the only thing they can do if they compromise the server is impersonate you on git. ...Exactly, that's arguably as bad as it gets. If you don't trust the jump box, do not use it. I had to use a "meh" jump box in a previous job. It was locked down to an extent, but a bunch of people could easily get root on it. Since I had to be active on that box for a bunch of work (not just "passing through" for the duration of a git pull) I set a very short lifetime on my ssh-add, so I got prompted for my passphrase whenever it expired and a key was requested. Not great, but better than leaving it active continually while I'm connected.