soft-serve vulnerable to SSRF via unvalidated LFS endpoint in repo import

QFinch assessment

An authenticated SSH user can force the server to make HTTP requests to internal/private IP addresses by running repo import with a crafted --lfs-endpoint URL. The initial batch request is blind (the response from a metadata endpoint won't parse as valid LFS JSON), but an attacker hosting a fake LFS server can chain this into full read access to internal services by returning download URLs that point at internal targets.

Potential impact

Any authenticated SSH user (any valid SSH key) can make the server send HTTP requests to arbitrary destinations, including internal networks and cloud metadata services. Concrete impact: - Port scanning / service discovery: Confirm reachability of internal hosts via timing and error responses - Cloud credential theft: Access cloud metadata endpoints (169.254.169.254) -- full credential extraction is possible through the fake-LFS-server chain unless IMDSv2 or equivalent is enforced - Internal API access: Read responses from internal services by routing LFS download URLs through the pipeline - Persistence: Mirror imports repeat the SSRF on every scheduled sync without further user action Reported by Vinayak Mishra GitHub: @vnykmshr

Technical context

The user-controlled endpoint flows through four files with zero validation: 1. User supplies the URL via --lfs-endpoint (pkg/ssh/cmd/import.go:20-41) The flag value is passed directly into proto.RepositoryOptions{LFSEndpoint: lfsEndpoint} at line 40 and then to be.ImportRepository(). 2. Access check passes for any authenticated user (pkg/ssh/cmd/cmd.go:172-187, pkg/backend/user.go:94-100) The import command uses checkIfCollab as its PersistentPreRunE. For a new repo name (which is normal during import -- you're creating it), AccessLevelForUser hits this path: This is by design -- any authenticated user can create repos via import or push (same model as Gitea/Gogs). The point isn't that the access control is wrong, just that any valid SSH key is enough to trigger the SSRF. 3. Endpoint flows to the LFS client unvalidated (pkg/backend/repo.go:170-194) When opts.LFSEndpoint is non-empty, it overrides the remote URL entirely. No URL validation, no IP check. It then flows through: lfs.NewEndpoint does URL parsing only -- no SSRF validation. lfs.NewClient calls newHTTPClient. 4. HTTP client has no protection (pkg/lfs/httpclient.go:24-31) Both the batch client and the BasicTransferAdapter use http.DefaultClient -- no SSRF protection, no IP validation, follows redirects. Compare with the webhook client that was added in v0.11.1: How the attack chains together: Stage 1 -- blind SSRF: The server sends a POST to <attacker-endpoint>/objects/batch (see httpclient.go:57). If the endpoint is a cloud metadata service like http://169.254.169.254/latest/meta-data/, the response won't be valid JSON, so the batch request fails with a parse error. The request is still sent though -- the attacker can confirm reachability via timing or error differentiation. Stage 2 -- reading internal responses via fake LFS server: If the attacker hosts a fake LFS server that returns valid batch responses, the BasicTransferAdapter follows the download URLs from the response: The l.Href field comes from the attacker's batch response. The a.client is the same unprotected http.DefaultClient. So the fake LFS server can point download URLs at internal targets like http://169.254.169.254/latest/api/token or http://10.0.0.1:8080/admin, and the response bodies get written to LFS object storage on disk. Since the attacker just created the repo and has read access, they can retrieve the stored objects through the normal LFS download API. Mirror sync persistence: When a repo is imported with --lfs-endpoint, the URL is persisted in the repo's git config at lfs.url (repo.go:175). If imported as a mirror (--mirror), the periodic sync job reads this config and uses the same unprotected LFS client: A single --mirror --lfs --lfs-endpoint <internal-url> import creates persistent SSRF that repeats on every mirror sync without further interaction. Two notes: - The batch request only fires if the imported repo contains LFS pointer blobs (checked via SearchPointerBlobs). The attacker needs to import a repo that has LFS objects -- easy to arrange with your own repo, but worth noting. - The import path in repo.go does not check the global cfg.LFS.Enabled flag -- it always processes LFS when the --lfs flag is passed. The mirror path (mirror.go:87) does gate on cfg.LFS.Enabled. So the import vector works regardless of server-level LFS configuration. Protection comparison: | Layer | Webhooks (v0.11.1+) | LFS import/mirror | |---|---|---| | URL validation at input | ValidateWebhookURL() | None | | Custom HTTP transport | secureHTTPClient with ValidateIPBeforeDial | http.DefaultClient | | Redirect blocking | CheckRedirect returns http.ErrUseLastResponse | Default (follows redirects) | | DNS rebinding protection | IP checked at dial time | None | Affected versions: - Introduced in v0.6.0 (commit ea6b9a4 added --lfs-endpoint flag) - Still present in v0.11.3+ (current main) - Not fixed by v0.11.1 webhook SSRF patch (GHSA-vwq2-jx9q-9h9f) -- that fix only covers pkg/webhook/, not pkg/lfs/ Suggested fix: The existing SSRF protections in pkg/webhook/validator.go and pkg/webhook/webhook.go are thorough and well-tested. The cleanest fix would be to extract them to a shared internal package and apply them to the LFS client: 1. Replace http.DefaultClient in pkg/lfs/httpclient.go with a secure client using ValidateIPBeforeDial in the transport and http.ErrUseLastResponse in CheckRedirect -- matching the webhook pattern. 2. Validate the endpoint URL in pkg/backend/repo.go (before lfs.NewEndpoint) and pkg/jobs/mirror.go (before creating the client) using the same checks ValidateWebhookURL performs. Both layers matter -- URL validation catches the obvious cases, ValidateIPBeforeDial at connection time catches DNS rebinding.