An unverified claim says the ChatGPT desktop app forwards your local SSH agent to remote machines, which would let an AI agent running there push code to your private GitHub repositories.
The claim comes from a forum post stating that the agent on a remote machine can authenticate with your local SSH credentials. Its one example is pushing to private repos on GitHub. The post offers no test or write-up of its own. It links only to a comment on pull request 28131 in OpenAI's Codex repository, and we have not confirmed that the comment shows what the post says it does. Nothing in the post is a statement from OpenAI.
Some background helps here. An SSH agent is a small background program that holds your private keys and signs login requests on your behalf. Agent forwarding, a long-standing OpenSSH feature, extends that to a remote machine. The keys never leave your laptop, but software on the far end can ask your agent to sign requests for as long as the connection stays open. Security guidance has long said to use it only with hosts you trust, since anyone who controls the remote session can use your keys the same way.
If the claim holds, what is new is who sits on the other end. Normally a person chooses to forward an agent to a specific server, typically with a single flag (ssh -A) on one connection. Here, an AI coding agent would hold that reach, and agents act on text they read, which sharpens the trust question.
The details that would decide the matter are whether the forwarding is on by default, whether users can switch it off, and whether it covers every key in the agent or only some. Until the Codex thread or OpenAI answers those, this is a pointer to a code comment, not a confirmed finding.