CVE-2026-112910

filelock: WindowsFileLock ignores mode: a 0o600 lock file and the ReadWriteLock database inherit the directory ACL

filelock on Windows ignores the mode argument. FileLock(path, mode=0o600) creates the same file as FileLock(path): one with no security descriptor of its own (the Windows record that says who may access a file). Its access list, or ACL, is whatever the lock directory hands down to new files. filelock creates the ReadWriteLock database, which it asks to be 0o600, the same way. In a directory whose inheritable ACL grants other local accounts access to new files, those accounts can open the lock file or database. They can take the byte-range lock (a lock on a span of bytes in a file) once the owner releases it and keep the owner's later acquire() calls waiting or timing out. They can read holder metadata an on_acquired hook left in a preserve_lock_file=True file. Where the directory hands down modify access, they can write to both files.

lock.mode still reports 0o600. The documentation neither promises an ACL on Windows nor says that Windows drops the mode. On Unix the same mode keeps other accounts out. Mutual exclusion between processes that follow the protocol holds on either platform. Affected versions are filelock 3.10.0 through 4.2.0 on Windows.

Anthropic's OSS Scanner reported this in private. Finding covered by this advisory:

• ANT-2026-0KEO0DDO [medium, CWE-732]: WindowsFileLock ignores mode; 0o600 lock files inherit the directory ACL.

Severity

CVSS 3.1 CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:U/C:L/I:L/A:L scores 5.3 (Medium).

AV:L (local): the other account needs a logon on the same host. PR:L (low privileges): an ordinary account is enough. UI:N (no user interaction). AC:L (low attack complexity): no race. C:L (partial confidentiality impact): the lock file (holder metadata) and the lock database are readable where 0o600 was meant to keep them private. I:L (partial integrity impact): in the directory layout the maintainers tested, the other account can also write both files, which lets it corrupt the lock's own state but nothing the lock protects. A:L (partial availability impact): the other account can hold the byte-range lock and make the owner wait or time out. S:U (scope unchanged).

Classification from the review: Windows permission failure with security impact only where callers rely on mode=0o600 across accounts. The review had marked it an unconfirmed Windows claim, since the scanner traced stand-in Win32 calls on Linux. The maintainers have since confirmed it on Windows with two local accounts (see Evidence). The scanner rated this medium; the maintainers keep medium.

Scope and preconditions

• The affected classes are FileLock and AsyncFileLock on Windows (WindowsFileLock), and the SQLite database of ReadWriteLock and AsyncReadWriteLock. Both go through _open_non_reparse_fd. The soft locks create their markers with os.open, which also keeps only a read-only bit of the mode on Windows; the scanner did not examine them further.
• The directory is one that other local accounts cannot write, which the documentation already requires on Windows, but whose inheritable entries grant them read (or modify) access to new files. Examples are a service's data directory readable by Users, or any directory with (OI)(CI) entries (entries that new files and sub-directories inherit) for another account. A directory whose ACL already excludes other accounts, such as one inside a user profile, hides the problem.
• The other account must open the file. According to the LockFileEx documentation, read access is enough to take the byte-range lock, so it can hold the lock whenever the owner does not. The lock file exists while the owner holds the lock, and stays after release with preserve_lock_file=True. The ReadWriteLock database is not unlinked at any point.
• Unix is not affected, because UnixFileLock creates the file with the mode and repeats it with fchmod. Mutual exclusion between cooperating processes is not affected.

Affected code

Scanned commit b831496b0273: _open_non_reparse_fd reduces the mode to one read_only bit. That bit is false for every accepted mode, because _check_mode rejects any mode without owner read and write, and an unset mode becomes 0o666. In _nt_open the OBJECT_ATTRIBUTES.SecurityDescriptor field stays unassigned, and read_only only chooses FILE_ATTRIBUTE_READONLY. _open_lock_database passes 0o600 through the same function on Windows.

Introduced

d0061e4 "added multiuser support and associated tests" (#192, 2023-03-15), first tag 3.10.0 (verified), added the mode parameter and on Windows handed it to os.open, which applies only the read-only attribute. The Windows backend has applied no more of the mode than that at any version, so this is a long-standing gap rather than a regression.

Mechanism

_open_non_reparse_fd(path, mode) computes read_only = not mode & _OWNER_WRITE and calls _nt_open(path, read_only=read_only). _nt_open fills an OBJECT_ATTRIBUTES with the NT path and OBJ_CASE_INSENSITIVE and leaves SecurityDescriptor NULL. It then calls NtCreateFile with FILE_OPEN_IF and FILE_ATTRIBUTE_NORMAL (or FILE_ATTRIBUTE_READONLY when read_only, which no accepted mode produces). Nothing consults the group and other bits. Per the NtCreateFile and OBJECT_ATTRIBUTES documentation, a NULL descriptor gives the new file default security, which on NTFS is the set of inheritable entries of the parent directory. A lock file created with mode=0o600 in a directory that hands down read access to Users is therefore readable by Users.

Nothing tells the caller. The constructor validates the mode, lock.mode returns 0o600, and has_explicit_mode is True. docs/concepts.rst says "filelock creates the file with it" and that "0o600 and 0o700 keep other UIDs out", with no Windows caveat. Only the _open_non_reparse_fd docstring and a comment in _soft.py record the limitation.

Reproduction

The scanner's reproducer for ANT-2026-0KEO0DDO, quoted as received. It ran against commit b831496b0273.

repro

#!/usr/bin/env python3
import os
import shutil
import subprocess
import sys
import tempfile
private_job = "/tmp/repro.d/job_with_private_lock.py"  # with FileLock(path, mode=0o600)
default_job = "/tmp/repro.d/job_with_default_lock.py"  # with FileLock(path)
on_windows_backend = "/tmp/repro.d/run_on_windows_backend.py"  # prints every Win32 call filelock makes
os.umask(0o022)
locks = tempfile.mkdtemp()
control: UnixFileLock, which leaves the lock file behind
subprocess.run([sys.executable, private_job, f"{locks}/private.lock"], check=True)
subprocess.run([sys.executable, default_job, f"{locks}/default.lock"], check=True)
unix_private_mode = os.stat(f"{locks}/private.lock").st_mode & 0o777
unix_default_mode = os.stat(f"{locks}/default.lock").st_mode & 0o777
trigger: the same two programs on WindowsFileLock, same lock name so the call lists can be compared whole
windows_private = [sys.executable, on_windows_backend, private_job, f"{locks}/job.lock"]
windows_default = [sys.executable, on_windows_backend, default_job, f"{locks}/job.lock"]
windows_private_calls = subprocess.run(windows_private, check=True, capture_output=True, text=True).stdout.splitlines()
windows_default_calls = subprocess.run(windows_default, check=True, capture_output=True, text=True).stdout.splitlines()
shutil.rmtree(locks)
check: mode=0o600 must reach the OS; on Windows that means a security descriptor on NtCreateFile
print(f"CONTROL  UnixFileLock FileLock(path, mode=0o600): lock file mode {unix_private_mode:#o}")
print(f"CONTROL  UnixFileLock FileLock(path):             lock file mode {unix_default_mode:#o}")
for call in windows_private_calls:
    print(f"OBSERVED WindowsFileLock FileLock(path, mode=0o600): {call}")
for call in windows_default_calls:
    print(f"OBSERVED WindowsFileLock FileLock(path):             {call}")
print(f"OBSERVED the two Win32 call lists are identical: {windows_private_calls == windows_default_calls}")
honoured_on_unix = unix_private_mode == 0o600 and unix_default_mode == 0o644
private_create = next(call for call in windows_private_calls if call.startswith("NtCreateFile("))
dropped_on_windows = (
    "ObjectAttributes.SecurityDescriptor=None" in private_create
    and windows_private_calls == windows_default_calls
)
print("BUG" if honoured_on_unix and dropped_on_windows else "OK")

repro.d/job_with_default_lock.py

import sys
from filelock import FileLock
with FileLock(sys.argv[1]):
    pass

repro.d/job_with_private_lock.py

import sys
from filelock import FileLock
with FileLock(sys.argv[1], mode=0o600):
    pass

repro.d/run_on_windows_backend.py

import ctypes
import fcntl
import os
import runpy
import sys
import types
Stand-ins for the ntdll, kernel32 and msvcrt functions filelock's Windows backend calls. Each prints the call with
the arguments filelock passed, then does the nearest Linux thing so the program runs to its end.
def RtlDosPathNameToNtPathName_U_WithStatus(dos_path, nt_name, _file_part, _relative_name):
    print(f"RtlDosPathNameToNtPathName_U_WithStatus({os.path.basename(dos_path)!r})")
    nt_name._obj.Buffer = os.path.abspath(dos_path)
    return 0  # STATUS_SUCCESS
def NtCreateFile(
    handle, access, attributes, _io_status, allocation_size, file_attributes, share, disposition, options, ea, ea_length
):
    object_attributes = attributes._obj
    print(
        f"NtCreateFile(DesiredAccess={access:#x},"
        f" ObjectAttributes.Attributes={object_attributes.Attributes:#x},"
        f" ObjectAttributes.SecurityDescriptor={object_attributes.SecurityDescriptor},"
        f" AllocationSize={allocation_size}, FileAttributes={file_attributes:#x}, ShareAccess={share:#x},"
        f" CreateDisposition={disposition}, CreateOptions={options:#x}, EaBuffer={ea}, EaLength={ea_length})"
    )
    path = object_attributes.ObjectName.contents.Buffer
    handle._obj.value = os.open(path, os.O_RDWR | os.O_CREAT | os.O_NOFOLLOW)
    return 0  # STATUS_SUCCESS
def RtlFreeUnicodeString(_nt_name):
    print("RtlFreeUnicodeString()")
def GetFileInformationByHandle(_handle, info):
    print("GetFileInformationByHandle()")
    info._obj.dwFileAttributes = 0x80  # FILE_ATTRIBUTE_NORMAL: not a reparse point
    return True
def LockFileEx(handle, flags, _reserved, bytes_low, bytes_high, _overlapped):
    print(f"LockFileEx(dwFlags={flags:#x}, nNumberOfBytesToLockLow={bytes_low}, nNumberOfBytesToLockHigh={bytes_high})")
    fcntl.flock(handle, fcntl.LOCK_EX)
    return True
def UnlockFileEx(handle, _reserved, bytes_low, bytes_high, _overlapped):
    print(f"UnlockFileEx(nNumberOfBytesToUnlockLow={bytes_low}, nNumberOfBytesToUnlockHigh={bytes_high})")
    fcntl.flock(handle, fcntl.LOCK_UN)
    return True
only a filelock that builds a security descriptor for the lock file calls this
def ConvertStringSecurityDescriptorToSecurityDescriptorW(sddl, _revision, descriptor, _size):
    print(f"ConvertStringSecurityDescriptorToSecurityDescriptorW({sddl!r})")
    descriptor._obj.value = 1  # any address but NULL
    return True
filelock sets argtypes on these at import; a lock taken without contention never calls them
def not_called(*_arguments):
    raise NotImplementedError
dlls = {
    "ntdll": types.SimpleNamespace(
        RtlDosPathNameToNtPathName_U_WithStatus=RtlDosPathNameToNtPathName_U_WithStatus,
        NtCreateFile=NtCreateFile,
        RtlFreeUnicodeString=RtlFreeUnicodeString,
        RtlNtStatusToDosError=not_called,
    ),
    "kernel32": types.SimpleNamespace(
        GetFileInformationByHandle=GetFileInformationByHandle,
        LockFileEx=LockFileEx,
        UnlockFileEx=UnlockFileEx,
        CloseHandle=not_called,
        OpenProcess=not_called,
        GetProcessTimes=not_called,
    ),
    "advapi32": types.SimpleNamespace(
        ConvertStringSecurityDescriptorToSecurityDescriptorW=ConvertStringSecurityDescriptorToSecurityDescriptorW,
    ),
}
Import filelock once on the real platform, so the standard library modules it uses load their Linux versions, then
drop it and import it again with the Windows branches active. A HANDLE is the Linux descriptor number.
import filelock  # noqa: E402
for name in [name for name in sys.modules if name.split(".")[0] == "filelock"]:
    del sys.modules[name]
ctypes.WinDLL = lambda name, use_last_error=False: dlls[name]
sys.modules["msvcrt"] = types.SimpleNamespace(open_osfhandle=lambda handle, _flags: handle, get_osfhandle=lambda fd: fd)
os.O_NOINHERIT = 0
sys.platform = "win32"
import filelock  # noqa: E402, F811
assert filelock.FileLock is filelock.WindowsFileLock
sys.argv = sys.argv[1:]
runpy.run_path(sys.argv[0], run_name="__main__")

Failure sequence

sequenceDiagram
    participant O as owner process: FileLock(path, mode=0o600)
    participant N as NtCreateFile
    participant F as lock file (inherits the directory ACL)
    participant U as other local account
    O->>N: _open_non_reparse_fd(path, 0o600): read_only=False, SecurityDescriptor=NULL
    rect rgb(255, 214, 214)
        N-->>F: created with the parent's inheritable entries only, group and other bits not consulted
        Note over F: the file reports mode 0o600 to the owner, the ACL says otherwise
    end
    O->>F: LockFileEx (holds the lock), release, unlock
    U->>F: open for read (and write, where the directory hands down modify)
    U->>F: LockFileEx through that handle
    O->>F: acquire() -> waits or Timeout while U holds the byte-range lock
    U->>F: read holder metadata left by on_acquired with preserve_lock_file=True

Evidence

The scanner's reproducer runs on a Linux image. It imports filelock's unmodified Windows backend with sys.platform == "win32" over stand-ins that print every Win32 call. The program with mode=0o600 and the one without produce identical call lists (RtlDosPathNameToNtPathName_U_WithStatus, NtCreateFile with FileAttributes=0x80 and ObjectAttributes.SecurityDescriptor=None, RtlFreeUnicodeString, GetFileInformationByHandle, LockFileEx, UnlockFileEx), while the Unix control produced 0o600 and 0o644 files. It prints OBSERVED the two Win32 call lists are identical: True and BUG. The scanner stated that it inferred the resulting ACL, and what another user can then do, from the documentation and did not run on Windows.

Maintainer triage: the stand-in reproducer prints BUG on the scanned commit b831496 and on the 4.2.0 release image.

Maintainer verification on Windows 11 with two local accounts: a script creates private.lock with FileLock(mode=0o600, preserve_lock_file=True), default.lock without a mode, and a ReadWriteLock database lock.db as the first account, in a directory granted (OI)(CI)(M) to a second account that is a member of Users. It then opens each file under the second account's logon token with LogonUserW and ImpersonateLoggedOnUser. On 4.2.0 the second account could open private.lock and lock.db for both read and write, the same as default.lock, and icacls listed only inherited (I) entries on all three. This confirms the claim the review had marked unconfirmed.

Impact

A program that passes mode=0o600 so that other accounts cannot open its lock file gets no such restriction on Windows. In a directory that hands down read access, another account can hold the lock and make the owner wait or time out, and read holder metadata in a preserved lock file. Where the directory hands down modify access, as in the maintainers' test, it can also write the lock file and the ReadWriteLock database. With the mode honoured, as on Unix, a user without write access to the directory can do none of this. The exposure ends there: the effect reaches only the lock file or lock database, and mutual exclusion between processes that follow the protocol still holds.

Workarounds

• On Windows, keep lock files and ReadWriteLock databases in a directory whose ACL already excludes other accounts (inside the user profile, or a directory with inheritance disabled and only the owner, SYSTEM and Administrators granted), rather than relying on mode.
• After creating a lock file with preserve_lock_file=True or a ReadWriteLock database, set its ACL yourself, for example icacls /inheritance:r /grant:r ":F" "SYSTEM:F" "Administrators:F"; a file that exists keeps its ACL on later opens.
• Do not store anything confidential in the lock file through on_acquired on Windows.

Patched version

4.2.1.

Fix: https://github.com/tox-dev/filelock/pull/835

Credits

Reported by Anthropic's OSS Scanner (automated analysis), 15 October 2026; reviewed by the maintainers.

• 2026-10-15 23:08 UTC: report received (36 findings; this one rated medium by the scanner)
• 2026-10-16 05:38 UTC: draft advisory created (first as part of GHSA-69hj-f68j-2879, then split out); two-account verification on Windows by the maintainers
• 2026-10-16 01:16 UTC: fix PR #835 opened