Archive entry names are attacker controlled. `rclone archive extract` stripped
only a leading `./` and then joined the entry name onto the destination
directory with `path.Join`, which collapses `..` segments. An entry such as
`../escaped.txt` extracted into `:s3:bucket/safe/prefix` therefore resolved to
`bucket/safe/escaped.txt`, outside the selected `prefix` directory - a path
traversal ("Zip Slip") attack that could create or overwrite sibling objects on
any destination remote.
Entry names are now validated before use: a leading `./` is still stripped (tar
archives created with `tar -czf archive.tar.gz .` rely on this), but any entry
with a `..` path component is rejected. Both `/` and `\` are treated as
separators when looking for `..`, as the local backend treats `\` as a path
separator on Windows.
Fixes: GHSA-4vr5-p2gc-h23p
S3 object keys are opaque names that may legally contain `..` segments. `serve
s3` built backend paths with `path.Join(bucket, key)`, which normalised the key
so a request such as `GET /bucket/../root-secret.txt` resolved to a file outside
the selected bucket elsewhere under the serve root. Listing prefixes and
multipart uploads were affected also.
This did not allow reading of files outside the root, but did allow reading of
files in the root which normally aren't visible; only directories are visible as
buckets normally.
Because `serve s3` maps keys to file paths it cannot represent every opaque S3
key, so rather than normalising keys (which would alias distinct keys onto one
file as well as allow traversal) it now rejects any key that is not already in
canonical path form - containing `..`, `.`, `//` or a leading or trailing slash
- with a 400 Bad Request, as MinIO does. Directory listing prefixes are
validated the same way but allow the empty bucket-root prefix and an optional
trailing slash.
Fixes: GHSA-8v25-v8p6-qf7v
When mounting or otherwise opening an S3 prefix without a trailing
slash, rclone probes the path with a HEAD request to see whether it is
actually a file. Since v1.72.0 (#8975) any error other than "not
found" from that probe was fatal, so credentials scoped to a prefix -
which return 403 rather than 404 for the prefix key - could no longer
open the prefix at all.
6440052fbd s3: fix single file copying behavior with low permission
A 403 on the probe is ambiguous: it can mean either "this is the file
you named but you may not HEAD it" or "this is a prefix you may list
but not HEAD". When the HEAD is not permitted we now fall back to a
listing to disambiguate: if the path has children it is treated as a
directory, otherwise it is treated as a file.
Fixes#9582
rclone's shared Google Drive and Google Photos client_id is being
retired and will stop working during 2026. When creating a new remote
that would use it, the config wizard now warns the user and asks the
user to enter their own client_id and secret instead. Service account
and environment auth are unaffected as they don't use the shared
client_id.
See: https://forum.rclone.org/t/google-drive-and-google-photos-users-action-required/54005
The shared Google Drive and Google Photos client_id is being retired and
will stop working during 2026. Warn users who rely on it (ie who have not
configured their own client_id) so they can create their own in advance.
The warning is only shown for auth flows that actually use the shared
client_id, not for service account, environment or anonymous auth.
See: https://forum.rclone.org/t/google-drive-and-google-photos-users-action-required/54005
The --disable-zip flag was registered manually and was missing from
OptionsInfo, so it could not be set over the rc interface. Move it
into OptionsInfo like serve webdav does, which keeps the command line
flag and also makes it settable via rc.
The mega backend keeps the whole account as an in-memory tree. go-mega's
Delete with destroy=true (hard_delete) removed the node from its lookup
table but left it in its parent's children, so a hard deleted file kept
showing up in directory listings until the tree was reloaded. In a
long-running mount or serve the file reappeared once the VFS directory
cache expired and re-read the backend.
Update go-mega to pick up the fix, and let the ghost test exercise
whichever delete mode the remote is configured with.
The mega backend keeps the whole account as an in-memory tree which
go-mega reconciles asynchronously from the server's event stream. After
an upload, delete or move the optimistic local update could be undone
moments later when go-mega replayed the corresponding server event,
re-adding a node to its parent. In a long-running process such as mount
or serve this could briefly show a just-deleted file in a listing, or
make a moved file reappear shortly afterwards.
Wait for the server to confirm uploads, deletes and moves so the
in-memory tree is settled before returning, matching what Rmdir and
Purge already do. The move wait was also previously started after the
server-side move so only the rename was covered.
Add an internal test reproducing the lingering listing via a tight
put/remove/list loop.
When a cached file was closed, --vfs-handle-caching kept its handle and
downloaders alive for a grace period and closed them later from a timer.
The deferred close drops the item lock while tearing down the
downloaders, leaving the file handle open. A reopen landing in that
window saw no grace timer and a live handle, failed to create the cache
file with "internal error: didn't Close file" and removed the cache
file, which hung the application reopening the file.
Reopens now wait for an in-progress grace-period close to finish so they
start from a fully closed item.
See: https://forum.rclone.org/t/opening-a-recently-closed-and-cached-file-hangs-rclone/53986/
A backend flag set on the command line to a value that happened to
equal its default was silently ignored, letting the config file win
instead.
For example --sftp-user defaults to the current user, so connecting as
that same user with --sftp-user=USER had no effect and caused rclone
to use the value from the config file.
See: https://forum.rclone.org/t/sftp-user-parsing-as-cli-argument-broken/53955
govulncheck reported four vulnerabilities in golang.org/x/image v0.41.0,
all reachable via the internxt backend's thumbnail/image upload path:
- CVE-2026-46604 (GO-2026-5066): panic decoding a TIFF image with an
out-of-bounds strip offset
- CVE-2026-46602 (GO-2026-5062): unbounded memory use from lack of a
limit on TIFF tile sizes
- CVE-2026-46601 (GO-2026-5061): panic on a WEBP VP8 alpha channel size
mismatch
- CVE-2026-33813 (GO-2026-4961): panic decoding a large WEBP image on
32-bit platforms
Updating to v0.43.0 fixes all four.
Before this change, with --files-from and --no-traverse, a single file
that could not be read (for example permission denied) stopped all the
other files in the list being copied.
This happened because the error was returned from the listing, which
caused the whole source listing to be discarded.
This change counts and logs such per-file errors and carries on, so the
readable files are still copied and rclone exits with a non-zero error
code.
Fixes#9115
Add a .github/CODEOWNERS file mapping backends and core subsystems to
their maintainers so GitHub auto-requests reviews. Move the per-area
responsibility mapping out of MAINTAINERS.md into CODEOWNERS, leaving
MAINTAINERS.md as the maintainer list with a pointer to CODEOWNERS.
Since v1.74.0, ncdu doubled every keystroke on Windows: pressing Enter
entered a directory and immediately descended a level deeper, "n" sorted
by name and then reversed the order, and so on. Arrow keys were
unaffected.
The cause is an upstream regression in github.com/gdamore/tcell/v2
pulled in by the v2.9.0 -> v2.13.8 bump in commit 23917555. The same
regression affects lazygit and micro.
There is no fixed tcell release, so pin tcell back to v2.9.0, the last
version before that commit, until the input handling is fixed upstream.
See: https://github.com/gdamore/tcell/issues/1124
See: https://github.com/jesseduffield/lazygit/issues/5344Fixes#9539
Convert the bug report and feature request templates to GitHub issue forms.
Rewrite the pull request template to require that non-trivial changes are
discussed in an issue first, and to spell out that backend changes need a
clean test_all run and a test account for the integration tester before
they can be merged. Document the latter requirement in CONTRIBUTING.md.
Remove the legacy ISSUE_TEMPLATE.md which is superseded by the forms.
The new "trace" dump flag attaches a net/http/httptrace ClientTrace to
each HTTP transaction and logs the connection level events - DNS
resolution, TCP connect, TLS handshake (including the negotiated TLS
version, cipher, ALPN protocol and server certificate), connection
reuse, request write and time to first response byte. Each line is
tagged with the time elapsed since the start of the transaction and the
request pointer so it can be correlated with the other dumps.
This is complementary to the existing dump flags: it shows how the
connection behaved rather than what was sent, which is useful for
debugging connectivity, DNS, TLS, proxy and keep-alive problems.
The new "errors" dump flag makes the HTTP dump conditional on the
transaction failing with a retryable error (a transport error, HTTP 429
or HTTP 5xx), so first-failure diagnostics can be captured without the
noise of dumping every transaction. The existing dump flags continue to
control what is dumped, for example --dump errors,bodies, and on its own
--dump errors dumps the headers.
TestS3Minio brings up a minio container via the fstest/testserver
framework, which exec's bash init.d scripts that shell out to docker.
This is not available on all platforms - Windows has no POSIX shell to
run the scripts, and macOS CI runners have no docker daemon - which
caused the build to fail there.
Add testy.SkipUnlessDocker to detect whether the framework can run and
skip the test when it cannot.
Previously serve s3 buffered every part of a multipart upload in memory
(in the gofakes3 S3 library) and concatenated them when the upload
completed, so memory use grew with the size of the upload.
serve s3 now streams the parts, in part-number order, into a single
PutStream upload to the underlying remote, which performs its own upload
with bounded memory. The whole file is never held in memory - memory use
is bounded by the parts in flight. This works for any remote that
supports PutStream (nearly all, including crypt) and for any part size,
so clients that don't produce uniform-sized parts (for example
PostgreSQL backup tools such as pgBarman and pgBackRest) work too.
Parts must arrive in ascending, contiguous part-number order; parts
uploaded out of order are buffered until their turn, and there is no
per-part retry (a failure aborts the whole upload). These trade-offs are
documented.
Passing --disable-multipart-streaming, or using a remote without
PutStream, reverts to buffering the parts in memory (the previous
behaviour); a one-off NOTICE is logged the first time this happens.
Fixes#7453
Run still uses a fresh local directory as the backing Fs that the
server wraps. RunWithBackend takes an extra remote name (e.g.
"TestS3Minio:") and uses a random subdirectory of that remote instead,
starting the matching fstest/testserver/init.d script on the way in
and tearing it down on the way out.
AuthProxy is only run for the local backend.
Before this change, if the user changed their password or public-key
and the auth proxy script returned updated config parameters for the
backend (eg a rotated api_key) rclone would continue to re-use the old
backend with the old config parameters out of the fscache.
This was because both the VFS cache and the fs/cache key were derived
from the user name only, so a change in the user's password or
public-key did not invalidate the cached backend.
Fix this by deriving the cache key from the user plus a hash of the
password/public-key, so a credential change forces a fresh backend.
The hash uses a per-process random HMAC key so the fragment that
appears in logs cannot be brute-forced offline.
ResetCounters unconditionally restarted the average loop, spawning a
ticker goroutine that pinned the StatsInfo even when no loop had been
running before. statsGroups.delete calls ResetCounters on every removed
group, so deleting N stats groups leaked N goroutines and prevented GC
of the underlying StatsInfo objects.
Only restart the loop if it was active before the reset.
A connection string can carry global.* options which change rclone's
process-wide configuration (e.g. global.http_proxy). This is
undesirable for the rc interface which was designed to have multiple
users or connections at once. The rc interface has the `_config`
mechanism for setting request scoped global config.
This blocks global.* options on all rc paths by marking the context as
a remote control request at the rc boundaries. fs.NewFs then skips
applying global.* to the process-wide config for a marked context.
The marker is reapplied in fs.CopyConfig, which is the call rclone
uses to detach context but keep config.
global.* options still apply to the individual backend they are set
on, exactly like override.* options; they just no longer leak into the
rest of the process. Remotes created directly on the command line are
unaffected as are remotes defined in the config file.
See: GHSA-qw24-gh76-8rvv
The --rc-serve GET/HEAD file serving path accepted bracketed inline
remotes from the URL and instantiated them, so a single
unauthenticated request could run a command as the rclone user via
backend options such as webdav bearer_token_command or sftp ssh, read
arbitrary local files, or change process-wide config via global.*
options.
This was the GET/HEAD equivalent of the POST hole fixed for
CVE-2026-41179, which only guarded the rc call dispatch path.
Now, unless the rc server has authentication configured or
--rc-no-auth is set, the serve path only allows remotes already
present in the config file: inline remotes, connection string
parameters and bare local paths are rejected. Connection string
global.* options are never honoured on the serve path, even when
authenticated.
See: GHSA-qw24-gh76-8rvv
- CVE-2026-42504: mime: quadratic complexity in WordDecoder.DecodeHeader
- CVE-2026-42507: net/textproto: arbitrary input are included in errors without any escaping
- CVE-2026-27145: crypto/x509: split candidate hostname only once
Accounts with shared-album zones (named CMM-*) returned these zones in
the photos changes/database listing alongside the real photo
libraries. rclone treated them as libraries and queried their albums,
but these zones have no CPLAlbumByPositionLive index so the query
failed with BAD_REQUEST / "Index has invalid data", aborting the whole
listing.
Only PrimarySync and SharedSync-* zones are photo libraries, so skip
any other zone during discovery.
See: https://forum.rclone.org/t/sync-to-icloud-gives-index-has-invalid-data-errors/53852
The build matrix relied on setup-go's built-in cache, which keys only on
go.sum with no job differentiation. All matrix jobs computed the same
cache key and raced to save it; since cache keys are immutable, only the
first job to finish saved its cache. That winner was usually a fast job
whose build cache contained none of the cross-compiled architectures, so
the compile_all and ci_beta steps started from a cold cache on every run.
Disable setup-go's cache and add two explicit actions/cache steps to the
build matrix, the android job and the lint job:
- the module cache (~/go/pkg/mod) depends only on go.sum, so it is
shared across all jobs under a single key; it used to be duplicated in
every job's cache. The downloaded module .zip archives are pruned
before saving as they are not needed to build from the extracted
module cache, roughly halving it to ~260 MiB per OS;
- the build cache (compiled artifacts) is specific to OS, arch and Go
version, so it is kept per job, keyed on the job name.
This lets the cross-compile steps reuse per-architecture build artifacts
and keeps the total cache within the repository limit.
Measured on CI, comparing a cold-cache run against the following
warm-cache run:
other_os 23m12s -> 3m35s (compile_all 14m -> 21s)
linux 23m13s -> 12m14s (deploy 11m -> 1m37s,
race test 8m -> 4m45s)
Both jobs now finish well under 15 minutes once the cache is warm.
JSON logs reported "source":"slog/logger.go:256" instead of the real
caller. getCaller skips logging-machinery frames by file path, but
release builds use -trimpath which rewrites the standard library slog
frame's path to "log/slog/logger.go" - matching neither the "/log/" nor
the "log.go" check, so it was reported as the source. Also skip frames
whose function belongs to the log/slog package, which is immune to
-trimpath.
Non-exportable Google documents such as Google My Maps, Sites and
Forms have no export format, so rclone silently leaves them out of all
listings. During a server side move/copy this means they are quietly
left behind.
Emit a one-off notice when such a document is skipped, telling the
user the files are invisible to rclone and pointing them at
--drive-show-all-gdocs to include them in server side copies and
moves.
When moving a file rclone removed the file from its old parent by
looking the parent ID up from the path using the directory cache. When
the source contained two directories with the same name and path, the
cache could resolve to the wrong duplicate, so the removeParents
request didn't match the file's real parent.
This left the file with both its old and new parent, which fails on
shared drives with:
A shared drive item must have exactly one parent., teamDrivesParentLimit
This uses the source object's actual parent ID instead when it is
known, falling back to the path lookup only when the object has zero
or multiple parents.
Listing the parts of a multipart upload with a MaxParts smaller than
the number of parts returned wrong PartNumber values and duplicate
parts on pages after the first, and the part number marker regressed
instead of advancing, causing clients to loop forever.
This bumps github.com/rclone/gofakes3 to v0.0.6 pick up the fix.
Drime's folder rename returns success immediately but the children
listing of the renamed folder briefly returns empty before settling,
which made VFS see an empty directory right after a successful rename
and broke subsequent operations on its contents.
After a successful DirMove, poll the renamed folder's listing until it
shows the pre-rename child count, with a 30s timeout and exponential
backoff. The backend integration test already tolerated this via its
own list-consistency retries; this brings VFS into line.
Fixes#9450
Large files (sent as multipart uploads) were placed in the wrong folder
for two reasons:
- the parent folder was sent as "parent_id", but the API ignores that
and expects "parentId", so the parent was never honoured
- relativePath was sent as the full path from the drive root, which made
the server build folders from it and silently drop any "0" path
segment (e.g. ".../data/0/file" lost the "0")
Send the parent as "parentId" and use just the leaf as relativePath,
matching the working single-part upload. This also lets us remove the
now-unneeded absolute-path resolution code.
Fixes#9392
Co-authored-by: Brian King <BrianDKing@gmail.com>
The drime origin returns a malformed response (reported by Cloudflare as
a 520 error) for a literal PUT request to the file-entries update
endpoint, which broke renaming, and so server-side copy and move.
Use a POST with the X-HTTP-Method-Override: PUT header instead - the API
routes this to the same handler and it works reliably.
Also retry Cloudflare 520-524 errors which may occur transiently.
When a SETSTAT request set the access and modification times, the
handler passed the modification time for both, discarding the
requested access time. Pass the requested access time through instead.
The VFS currently ignores the atime, but it might use it one day.
The statvfs@openssh.com extension was advertised but returned an
unsupported status, so clients couldn't query the amount of free and
used space. Implement it using the VFS Statfs method, which reports the
backend's usage where the backend supports About.
The SFTP serve handler ignored the size attribute of SETSTAT/FSETSTAT
requests, only acting on the modification time. This meant a client
asking to truncate a file (eg setting the final size of an upload, or
an explicit truncate) had no effect at all.
This respects the size attribute (if present) by truncating the file
to the requested size.
The SFTP serve write handler always opened files with O_TRUNC,
ignoring the flags requested in the SFTP OPEN packet. Some clients
(notably WinSCP's "Process in Background", which resumes an upload on
a second connection) re-open the partially written file without the
truncate flag and continue writing from the offset they had reached,
relying on the existing data being preserved. Forcing O_TRUNC zeroed
that prefix, so the start of the uploaded file ended up as a block of
zero bytes.
This fix respects the requested open flags instead so a resume open
without truncate keeps the already written data intact.
See: https://forum.rclone.org/t/rclone-serve-sftp-winscp-background-mode-uploading-causes-file-corruption/53841
Previously the Mount RPC ignored the path component of the mount
request, so `server:/sub/dir` and `server:/` both landed at the root
of the served remote. The Mount handler now cleans the requested path
with path.Clean, looks it up in the VFS and serves a billy.Filesystem
rooted at that directory, refusing the mount if the path does not
exist or is not a plain directory.
A pathRewriter cache wraps the inner handle cache so that the same
file always produces the same NFS file handle regardless of which
mount minted it (and stable across server restarts for the disk and
symlink caches). This matches the traditional NFS expectation that a
subpath mount behaves like `cd` into a subtree.
nfsmount gains a --nfs-mount-path flag (default /) so clients can
select a subpath at mount time. This replaces a latent misuse of
--volname as the NFS mount path that was previously masked by the
server ignoring it.
Fixes#9442
The golang.org/x/net/http2/h2c package was deprecated in v0.54.0 in
favour of setting the http.Server Protocols field to enable unencrypted
HTTP/2.
This replaces the h2c.NewHandler wrapping added in e863f751f with
http.Server.Protocols, which is supported by the standard library
since Go 1.24.
Note that the stdlib only supports HTTP/2 prior-knowledge on cleartext
connections, not HTTP/1.1 Upgrade: h2c negotiation. In practice clients
use prior-knowledge or require TLS, so this should not affect users.
Upgrade to v0.55.0 of golang.org/x/net in order to address:
- CVE-2026-42506: html: incorrect handling of namespaced elements in foreign content
- CVE-2026-39821: idna: failure to reject ASCII-only Punycode-encoded labels
- CVE-2026-42502: html: incorrect handling of HTML elements in foreign content
- CVE-2026-25680: html: denial of service when parsing arbitrary HTML
- CVE-2026-25681: html: incorrect handling of character references in DOCTYPE nodes
- CVE-2026-27136: html: duplicate attributes can cause XSS
Upgrade to version v0.52.0 of golang.org/x/crypto to address:
- CVE-2026-46598: ssh/agent: pathological inputs can lead to client panic
- CVE-2026-46597: ssh: byte arithmetic causes underflow and panic
- CVE-2026-39828: ssh: bypass of certificate restrictions
- CVE-2026-39835: ssh: server panic during CheckHostKey/Authenticate
- CVE-2026-39833: ssh/agent: key constraints not enforced
- CVE-2026-39832: ssh/agent: agent constraints dropped when forwarding keys
- CVE-2026-39827: ssh: memory leak when rejecting channels can lead to DoS
- CVE-2026-39830: ssh: client can cause server deadlock on unexpected responses
- CVE-2026-39829: ssh: pathological RSA/DSA parameters may cause DoS
- CVE-2026-39831: ssh: bypass of FIDO/U2F security keys physical interaction
- CVE-2026-39834: ssh: infinite loop on large channel writes
- CVE-2026-42508: ssh/knownhosts: auth bypass via unenforced @revoked status
- CVE-2026-46595: ssh: VerifiedPublicKeyCallback permissions skip enforcement
Upgrade to version v0.41.0 of golang.org/x/image to address:
- CVE-2026-42500: bmp: panic when reading out of bound palette index
- CVE-2026-33809: tiff: excessive resource consumption in PackBits decompression
Upgrade to version v0.45.0 of golang.org/x/sys to address:
- CVE-2026-39824: windows: integer overflow in NewNTUnicodeString
MEGA S4 has launched a new dedicated domain megas4.com with three new
endpoints, including their first Asia-Pacific location in Tokyo, plus
new endpoints in Paris and Barcelona. The existing Amsterdam,
Luxembourg, Montreal and Vancouver locations have been renamed onto
the new domain.
All current s4.mega.io endpoints continue to be supported and are
retained as legacy options. New accounts use the megas4.com endpoints.
The "Import/Export of google documents" section in drive.md and the
"--b2-versions" examples section in b2.md were both at H4 instead of H3,
which excluded them from the ToC even though they are top-level topics
in their respective sections.
The metrics_addr option was registered twice: once explicitly and once
implicitly via AddPrefix(libhttp.ConfigInfo, "metrics", ...). Both
pointed at the same MetricsHTTP.ListenAddr field, so options/info
returned a duplicate entry.
Drop the explicit entry and use SetDefault to keep the empty default
(so the metrics server stays off unless configured), matching the
pattern already used for rc_addr.
Fixes#9419
- CVE-2026-42501: cmd/go: malicious module proxy can bypass checksum database
- CVE-2026-39825: net/http/httputil: ReverseProxy forwards queries with more than urlmaxqueryparams parameters
- CVE-2026-39836: net: panic in Dial and LookupPort when handling NUL byte on Windows
- CVE-2026-42499: net/mail: quadratic string concatenation in consumePhrase
- CVE-2026-39820: net/mail: quadratic string concatentation in consumeComment
- CVE-2026-39819: cmd/go: "go bug" follows symlinks in predictable temporary filenames
- CVE-2026-39817: cmd/go: "go tool pack" does not sanitize output paths
- CVE-2026-33814: net/http: infinite loop in HTTP/2 transport when given bad SETTINGS_MAX_FRAME_SIZE
- CVE-2026-39826: html/template: escaper bypass leads to XSS
- CVE-2026-33811: net: crash when handling long CNAME response
- CVE-2026-39823: html/template: bypass of meta content URL escaping causes XSS
The bisync normalization test relies on uploading distinct NFC and NFD
versions of the same filename and on the backend supporting in-place
modtime updates. Dropbox normalizes unicode server-side (NFD -> NFC)
and can't set modtime in place, so the test inevitably takes a
different code path on Dropbox and the log diverges from the golden
output without any functional difference.
operations.NeedTransfer's equality check may have deleted pair.Dst as
a precursor to re-uploading it if SetModTime returns
ErrorCantSetModTimeWithoutDelete (e.g. Dropbox). If so skip the eager
delete of the destination if --fix-case will rename it to a different
name. The rename itself replaces the destination, and any subsequent
re-upload happens at the correctly-cased path.
See: #8881
Nick Craig-Wood