SeaweedFS is a distributed storage system for files and blobs. In versions prior to 4.40, an authenticated S3 principal with permissions scoped to a nested object key can overwrite a different object outside that scope by calling PutObjectAcl on the key it is allowed to access. The handler authorizes the request against the requested nested key but then writes the updated entry back to the bucket root rather than the key's actual parent directory, so an ACL change on allowed/protected.txt is instead applied to protected.txt at the bucket root. Because the update carries the full entry rather than only ACL metadata, an existing target object is overwritten with the content, metadata, owner information, and ACL of the scoped object, bypassing the object-level action scoping configured through the static S3 identity file. This issue is fixed in version 4.40.

SeaweedFS is a distributed storage system for files and blobs. In versions from 3.88 through 4.39, the SFTP server evaluates configured path permissions with a literal string-prefix comparison, so a user scoped to a path is also granted the same access to any sibling path whose name merely begins with the same characters. A user granted access to /tenants/alice therefore also matches /tenants/alice-archive, /tenants/alice2, and similar siblings, because the check does not require a path-component boundary. An authenticated low-privilege SFTP user with a root home directory and narrow path permissions can thereby cross the configured ACL boundary to read another tenant's files, and to overwrite them if granted write, all through the documented SFTP service with its own valid credentials. This issue is fixed in version 4.40.

SeaweedFS is a distributed storage system for files and blobs. In versions 4.39 and earlier, the S3 API accepts an external OIDC JWT sent directly in the Authorization header and maps it to an IAM role without enforcing that role's trust policy, so a federated user can assume a role they are not permitted to hold. The standard STS AssumeRoleWithWebIdentity path rejects such a token when the role's trust policy does not trust the token's federated provider, but the direct S3 bearer path validates only the token itself and then authenticates as the mapped role and evaluates that role's attached S3 permissions. As a result, a valid OIDC user whose token would be denied the role through STS can obtain the role's S3 access, including object read, write, and delete, by presenting the raw OIDC JWT directly to the S3 API. This issue is fixed in version 4.40

SeaweedFS is a distributed storage system for files and blobs. In version 4.39, the filer's TUS resumable-upload handler checks JWT allowedprefixes scoping only when a session is created, letting a low-privilege tenant hijack another tenant's upload session to write content to filer paths their own token forbids. The HEAD, PATCH, and DELETE verbs that act on an existing session by its id never verify that the session's stored target path falls within the caller's allowed prefixes, so a tenant who obtains another upload's session identifier can PATCH attacker bytes into it and, on completion, have the file land at the victim's out-of-scope path. The same token can also DELETE other tenants' sessions and HEAD them to read upload progress and size, defeating the JWT prefix isolation. This vulnerability only affects deployments that configure filer JWT signing and have TUS uploads enabled. This issue is fixed in version 4.40.

SeaweedFS is a distributed storage system. Prior to 4.24, the weed/server/filerserverhandlers.go allowedprefixes authorization check used strings.HasPrefix on raw path strings, so a filer JWT scoped to /tenant1 also authorized sibling paths such as /tenant1234, /tenant1-old, and /tenant1backup, enabling cross-tenant reads and writes with a valid scoped token. This issue is fixed in version 4.24.

SeaweedFS is a distributed storage system. Prior to 4.24, the filer registers the SeaweedIdentityAccessManagement gRPC service without mandatory authentication when jwt.filersigning.key is unset, allowing any client that can reach the filer gRPC port to invoke CreateUser, CreateAccessKey, PutPolicy, and related IAM RPCs to mint credentials and gain S3 administrative control. This issue is fixed in versions 4.24.

SeaweedFS is a distributed storage system. In versions 4.08 through 4.33, requests signed with SigV4 service s3tables are routed to the S3Tables management API where authorization collapses account-less S3 identities into the shared admin account and fails open, allowing an authenticated low-privileged S3 user to enumerate administrator-owned table bucket names and ARNs. This issue is fixed in version 4.34.

SeaweedFS is a distributed storage system. Prior to 4.34, the S3 API gateway does not reject dot-dot path segments in the X-Amz-Copy-Source header used by CopyObject and UploadPartCopy, allowing an authenticated identity scoped to one bucket to read objects from other buckets through server-side copy. This issue is fixed in version 4.34.

SeaweedFS before 4.34 contains a path traversal vulnerability in the S3 gateway DeleteMultipleObjectsHandler that allows authenticated S3 principals with write access to a single bucket to delete arbitrary objects in other tenants' buckets by supplying object keys containing ../ sequences in the DeleteObjects XML request body. Attackers can bypass authorization controls through a confused deputy condition, as the validateRequestPath middleware only inspects URL-captured path variables and never examines request-body keys, allowing the filer path to collapse directory traversal sequences and resolve deletions outside the authorized bucket.

SeaweedFS before 4.30 reflects the callback query parameter verbatim into responses served with Content-Type application/javascript in the shared writeJson helper (weed/server/common.go), with no callback-name validation, no X-Content-Type-Options: nosniff header, and no CORS allow-list. Every JSON endpoint that uses writeJson - including the unauthenticated master endpoints /dir/status, /dir/lookup and /cluster/status, the volume server /status, and the filer directory listing, all reachable in the default configuration (no -whiteList, no security.toml, bound to 0.0.0.0) - can therefore be loaded cross-origin via a script tag with a chosen callback, letting a third-party web page read cluster topology, volume server URLs and gRPC ports, file identifiers, and directory listings. Because the callback string is reflected at the start of the body and no nosniff header is sent, MIME-sniffing clients may also interpret the reflected content as HTML.

Summary

The S3 API gateway and the Iceberg REST catalog gateway construct their routers with mux.NewRouter().SkipClean(true). With path cleaning disabled, a .. segment inside the URL survives routing, so a request such as:

GET /bucket-A/../evil-bucket/key

is matched as bucket=bucket-A, object=../evil-bucket/key. The captured object key is then joined into a filer path with util.JoinPath (S3) / path.Join (Iceberg), which collapse the .. server-side, so the actual read or write lands in evil-bucket.

The captured path variables were never validated for traversal segments before reaching the handlers, so bucket isolation depended on downstream checks rather than on the path itself.

Impact

- With authentication disabled (enableAuth=false): direct cross-bucket read and write. An object key containing .. resolves to and operates on a different bucket than the one named in the request path. - With authentication enabled (enableAuth=true): an authorization confused-deputy. IAM evaluates the policy against the mux {bucket} variable (bucket-A) in iam.authRequestWithAuthType, while the I/O is performed against the traversed target (evil-bucket). A principal authorized for one bucket can therefore reach objects in another bucket it has no grant for. This breaks tenant isolation.

The same class of traversal applies to the Iceberg REST catalog's {prefix}, {namespace}, and {table} path variables.

%2e%2e-encoded and ..\ (backslash) variants are equivalent, because gorilla/mux URL-decodes captured variables and NormalizeObjectKey folds \ to / before the path is used.

Affected components

- S3 API gateway (weed s3, and the S3 endpoint embedded in weed server) - Iceberg REST catalog gateway

Affected versions

All releases prior to 4.30.

Patched version

4.30 and later.

Proof of concept

With a bucket evil-bucket containing secret.txt, and a caller that only has (or needs no) access to bucket-A:

GET /bucket-A/../evil-bucket/secret.txt HTTP/1.1 Host: <gateway>

The response returns the contents of evil-bucket/secret.txt. The encoded form GET /bucket-A/%2e%2e/evil-bucket/secret.txt behaves identically.

Remediation

Upgrade to SeaweedFS 4.30 or later. The fix adds a validation middleware to both gateway routers that rejects any captured path variable containing a . or .. segment, a NUL byte, an embedded slash/backslash in single-segment slots, or an empty captured value, before any handler runs.

Workarounds

For deployments that cannot upgrade immediately, place a reverse proxy in front of the gateway that normalizes the request path and rejects requests whose path contains .., %2e%2e, or backslash sequences. Note that disabling auth removes the only remaining barrier, so do not rely on enableAuth=false deployments being protected by anything.

Resources

- Fix: https://github.com/seaweedfs/seaweedfs/pull/9687 (commit dd1b428)

Credits

Reported responsibly by Denis Abashkin (@dadbravo).