CVE-2026-112941: 
Python 취약성 분석 및 완화

CVE-2026-112941 is a Time-of-Check Time-of-Use (TOCTOU) race condition vulnerability in the filelock Python library (maintained by tox-dev) affecting SoftReadWriteLock and AsyncSoftReadWriteLock. Versions 4.1.0 through 4.2.0 are vulnerable; the issue was introduced in version 4.1.0 and fixed in 4.2.1. The vulnerability was reported by Anthropic's OSS Scanner on October 15, 2026, and publicly disclosed on October 17, 2026. It carries a CVSS v3.1 base score of 3.3 (Low) (GitHub Advisory, Red Hat CVE).

The root cause is a CWE-367 (Time-of-Check Time-of-Use) race condition in OsFiles.prepare(), which checks once per lock object (using a _prepared boolean shortcut introduced in commit 367b111) that .rw, .rw/gen, and .rw/holders are real directories and not symbolic links. However, every subsequent record operation (create, link, replace, unlink, listdir) resolves the full path again by name, and O_NOFOLLOW only protects the final path component — not intermediate directory components. A local attacker who can write to the directory containing the lock (e.g., a shared 0777 directory without the sticky bit, or a sticky directory like /tmp where the attacker created .rw first) can rename the legitimate .rw tree aside and plant a replacement tree whose gen or holders subdirectory is a symbolic link pointing to an arbitrary directory writable by the lock user. Because the live lock object never re-runs the directory check, subsequent lock acquisitions by the victim process transparently write lock records (epoch files, generation records, and HEAD renames) into the attacker-chosen target directory. There is no timing race to win — the swap occurs between two acquisitions of a long-lived lock object (GitHub Advisory, Fix PR #842).

A local low-privileged attacker can cause the lock user's process to write filelock-controlled files (epoch records, generation records, and a renamed HEAD file) into any directory the lock user can write — including private 0700 directories the attacker cannot directly access. The bytes written are filelock's own protocol records (not attacker-chosen content), but the effect is destructive: for example, the attacker can corrupt a git repository's .git/HEAD, rendering it unusable (git status reports fatal: not a git repository). Additionally, the sweep mechanism can delete files with 32-hex-digit names in the target directory, and generation log compaction can delete 20-digit filenames. There is no confidentiality impact and no direct availability impact on the lock itself, but integrity of targeted files is compromised (GitHub Advisory).

There is no public proof-of-concept exploit and no evidence of in-the-wild exploitation as of the disclosure date (Feedly). The EPSS score is 0.0, and the vulnerability is not listed in the CISA Known Exploited Vulnerabilities catalog. Exploitation requires a local account on the same host and write access to the directory containing the lock path (or the ability to create .rw before the lock user in a sticky directory). The attack has no timing race — the swap simply needs to occur between two acquisitions of a long-lived lock object, making it straightforward for a patient local attacker (GitHub Advisory).

1. Identify a target: Locate a system where a victim process uses SoftReadWriteLock or AsyncSoftReadWriteLock (filelock 4.1.0–4.2.0) with a lock path in a directory the attacker can write to (e.g., a 0777 non-sticky shared directory like /srv/shared, or a sticky /tmp where the attacker can create .rw first).
2. Wait for first acquisition: Allow the victim's long-lived lock object to complete its first write_lock() acquisition, which creates app.lock.rw (owned by the victim, mode 0700) and sets _prepared = True.
3. Swap the lock tree (non-sticky parent): As the attacker, rename the victim's app.lock.rw aside (rename(app.lock.rw, app.lock.rw.aside)), then create a new app.lock.rw directory with a world-writable holders/ subdirectory and a gen entry that is a symbolic link pointing to the target directory (e.g., /home/victim/project/.git).
4. Trigger second acquisition: Wait for the victim's process to acquire the lock a second time. Because _prepared is True, prepare() returns immediately without re-checking the tree.
5. Observe file writes into target: The victim's lock operations now traverse the symlink: _incarnation() creates epoch in the target, commit() creates generation records and renames a record over HEAD in the target directory — all with the victim's privileges, even though the attacker cannot directly access the target directory.
6. Verify impact: The target directory (e.g., .git) now contains filelock protocol files, and HEAD has been overwritten with a filelock record, breaking the git repository (GitHub Advisory).

• File System: Unexpected files named epoch, 20-digit numeric names (e.g., 00000000000000000001), or .commit-* temporary files appearing in directories unrelated to the lock path (e.g., inside a .git directory or other application data directories).
• File System: A file named HEAD in a git repository or other sensitive directory whose content has been replaced with a filelock record (e.g., content matching \nfilelock-rw/1\ngeneration=N\n).
• File System: Presence of a <lockname>.rw.aside directory alongside the expected lock tree, indicating the original lock tree was renamed by an attacker.
• File System: A <lockname>.rw/gen entry that is a symbolic link (detectable via ls -la or lstat) rather than a real directory, pointing to a sensitive directory.
• Process: Unexpected file creation or rename operations by the lock-user process in directories outside the expected lock tree, observable via auditd or inotifywait monitoring on sensitive directories.

Upgrade filelock to version 4.2.1 or later, which fixes the vulnerability by opening .rw, gen, and holders via directory file descriptors (O_NOFOLLOW) for every record operation on platforms supporting dir_fd, and by repeating ownership checks whenever a replaced tree is detected (filelock 4.2.1 Release, Fix PR #842). If immediate upgrade is not possible, apply these workarounds: keep the lock path in a directory owned exclusively by the locking account (or root) with no group or other write bits; avoid placing SoftReadWriteLock paths in world-writable directories like /tmp unless the locking account creates .rw first with mode 0700; and prefer creating a fresh lock object per unit of work (rather than a long-lived singleton) so that prepare() re-runs its directory check on each use (GitHub Advisory).

The vulnerability was reported privately by Anthropic's OSS Scanner (automated analysis) on October 15, 2026, under internal identifier ANT-2026-WXXRI4M4, and the maintainer (gaborbernat) opened a fix PR within hours of receiving the report. The advisory notes that Anthropic's scanner rated the issue as high severity, while the maintainers scored it as low (CVSS 3.3) based on the limited integrity impact (filelock's own records, not attacker-chosen bytes). No significant broader community or media reactions have been identified beyond the GitHub advisory and Red Hat tracking (GitHub Advisory, Red Hat CVE).