NLTK: WekaClassifier passes an unvalidated caller-controlled model path to the weka JVM, bypassing nltk.pathsec containment

Summary

WekaClassifier in NLTK passes a caller-controlled model file path to the weka Java VM without any nltk.pathsec validation. WekaClassifier.train puts the caller's model_filename verbatim into the weka argv ["-d", model_filename, ...] (weka writes the trained model there), and WekaClassifier._classify_many puts self._model into ["-l", self._model, ...] (weka reads it). Under nltk.pathsec.ENFORCE=True (the default on current HEAD), NLTK's stated invariant — established by advisory CVE-2026-81726 / GHSA-8mgp-746c-j5xp — is that caller-controlled model-artifact paths must be bounded to the allowed NLTK data roots before any native tool opens them. WekaClassifier is the one java()-based tool wrapper that skips this entirely, giving an untrusted workflow arbitrary-location file write (-d) and an arbitrary file read/parse oracle (-l), defeating pathsec containment.

This is a standalone, unfixed variant of the same class as CVE-2026-81726, on a code path the CVE-2026-81726 fix never touched. That advisory enumerates 6 functions (TransitionParser.train, TransitionParser.parse, AveragedPerceptron.save, AveragedPerceptron.load, PerceptronTagger.save_to_json, save_maxent_params) — WekaClassifier is not among them, and weka.py contains no validate_tool_path/validate_model_resource at all.

Root Cause

WekaClassifier.train(cls, model_filename, ...) builds cmd = [javaclass, "-d", model_filename, "-t", train_filename] (nltk/classify/weka.py:256) and WekaClassifier._classify_many builds [..., "-l", self._model, "-T", <staged>, ...] (weka.py:131-133); self._model is stored unvalidated in __init__ (weka.py:108-110). Only the internally-generated train.arff/test.arff are staged via make_staging_dir; the model path is passed as-is. nltk.internals.java() (internals.py:388-431) validates the classpath (_verify_jar_sandbox), the JVM launcher options (_validate_java_options, the CVE-2026-12841 fix), the main-class token cmd[0], and rejects @argfile tokens in any position — but it explicitly permits program arguments starting with - and never inspects the interior of a program-arg path value. So -d <path> / -l <path> flow unchecked to subprocess.Popen(final_cmd, shell=False) and the weka JVM performs an unbounded native open/write on the caller-chosen path.

Sibling wrappers prove the intended pattern: CRFTagger.train/set_model_file call validate_tool_path (crf.py:134,300); MaltParser.generate_malt_command validates -i/-o via validate_tool_path (malt.py:340,344), its model via validate_model_resource (malt.py:326) and working_dir via validate_path (malt.py:206). weka.py already uses pathsec elsewhere (validate_path for WEKAHOME, pathsec_open for ARFF writes) but never for the model path.

Impact

Under pathsec.ENFORCE=True (default on HEAD), an untrusted workflow that chooses the weka model save/load path obtains:

• Arbitrary-location file write via -d — overwrite or create a file at an attacker-chosen absolute or ..-traversal path outside the allowed NLTK data roots (content constrained to weka's serialized-model format; the target choice is the integrity impact). This corrupts/clobbers arbitrary chosen files.
• Arbitrary file read/parse oracle via -l — weka loads an out-of-root path as a model, yielding an existence/parse differential (file content is not echoed back → confidentiality impact is limited).

This defeats the pathsec containment advisory CVE-2026-81726 promises, at the same privilege level / threat model that advisory was accepted under.

Proof of Concept

import nltk.pathsec as ps; ps.ENFORCE = True
from nltk.classify.weka import WekaClassifier
# write-side: model written OUTSIDE any allowed NLTK data root (refused for CRFTagger, allowed here)
WekaClassifier.train("/tmp/outside_sandbox/evil.model",
                     [({"a": 1}, "pos"), ({"a": 0}, "neg")])
# read-side: -l opens an arbitrary path
clf = WekaClassifier(fmt, "/home/victim/secret.model")
clf.classify_many([{"a": 1}])

The argv construction — the vulnerable step — is confirmed without weka installed (see Bypass Evidence); a configured weka.jar is needed only to observe the JVM's actual file I/O.

Attack Chain

1. Entry. Action: an app running with pathsec.ENFORCE=True lets an untrusted workflow choose the weka model path and calls WekaClassifier.train("../../../home/victim/x", featuresets) (write) or WekaClassifier(fmt, "/home/victim/secret.model").classify_many(...) (read). Guard: none — plain classmethod/constructor parameter (weka.py:230,108-110). Bypass proof: grep shows weka.py uses no validate_tool_path/validate_model_resource/validate_tool_dir (NONE).
2. Command build. Action: the path is placed into cmd = [javaclass, "-d", model_filename, "-t", train_filename] (weka.py:256) / ["-l", self._model, ...]. Guard: only the internally-generated train.arff is staged via make_staging_dir; the model path is passed as-is. Bypass proof: runtime capture yielded ['weka.classifiers.bayes.NaiveBayes', '-d', '/tmp/outside_sandbox_weka.model', '-t', '<staged>/train.arff'] — the raw out-of-root path is cmd[2].
3. java() sink. Action: java(cmd, classpath=_weka_classpath, ...) → subprocess.Popen(final_cmd) (internals.py:431). Guard: java() validates classpath, JVM options, cmd[0], and rejects @argfile tokens. Bypass proof: none of these inspect the -d/-l program-arg value; internals.py explicitly documents that program args after the main class may start with -, and the path is neither cmd[0] nor an @-token, so every java() gate passes. Negative control: the same path through validate_tool_path(..., for_write=True, must_exist=False) raises PermissionError.
4. Impact. Action: the weka JVM opens the path outside the allowed roots → arbitrary-location write (-d) / parse-existence oracle (-l). Guard: pathsec is never consulted (no validate_tool_path, no pathsec_open). Bypass proof: the directly-analogous sibling CRFTagger.train calls validate_tool_path(model_file, must_exist=False) (crf.py:300); weka does not, and commit ee6d6138 hardened weka's other sinks (scratch files, WEKAHOME jar) while leaving the model path unvalidated.

Bypass Evidence

• Negative control (guard is real): validate_tool_path("/tmp/outside_sandbox_weka.model", for_write=True, must_exist=False) → PermissionError; validate_tool_path("../../../home/victim/.ssh/authorized_keys", ...) → PermissionError.
• Sink reached unvalidated (guard absent): captured write-side argv ['weka.classifiers.bayes.NaiveBayes', '-d', '/tmp/outside_sandbox_weka.model', '-t', '<staged>/train.arff'] and read-side [..., '-l', '/home/victim/secret.model', '-T', '<staged>/test.arff', '-p', '0'].
• Version: git show v3.10.3:nltk/classify/weka.py contains both the -d model-path build (line 235) and the -l read, and NO validate_tool_path.

Affected Versions

<= 3.10.3 (present on the latest release tag v3.10.3 and on HEAD c06cd264). No published fix has ever targeted the weka -d/-l model path; commits after v3.10.3 (ee6d6138, 92aeada6, a9752559) touched weka.py for other sinks but did not add model-path validation.

Suggested Fix

Before building the -d/-l/-T command, validate the model path with pathsec, mirroring CRFTagger/MaltParser:

validate_tool_path(model_filename, context="WekaClassifier.train", for_write=True, must_exist=False)
validate_tool_path(self._model, context="WekaClassifier._classify_many")

Reported by zx (Jace) — GitHub: @manus-use

Fix status (2026-09-28)

Status: MERGED on develop (unreleased; v3.10.3 of 2026-08-12 predates the fix).

PR: #3863 "Bound the WekaClassifier model path to the pathsec data roots", merged 2026-09-06, merge commit bc2878725 on develop. Fix commit: 6a1760629.

Probe: nltk/test/unit/security_probes/ghsa_j456_xh4h_cpf2.py, added by PR #3905 "Add GHSA-j456 live probe; expand the GHSA-8mgp probe attack set" (commit a96715264, merged 2026-09-20, merge commit 84d987bf8); verdict on develop 2fa3f48fe: FIXED (WekaClassifier -l and -d refuse out-of-root model paths).

Release: none yet; no release tag contains 6a1760629.