CVE-2026-128136: NLTK: AllowlistUnpickler does not deny numpy submodule-local file-I/O sinks
NLTK (Natural Language Toolkit) is a suite of open source Python modules, data sets, and tutorials supporting research and development in Natural Language Processing. From 3.10.0 until 3.10.4, allowlistedpickleload keys dangerous NumPy globals by export name, allowing submodule-local numpy.lib file input/output callables to evade the denylist under the broad module allowance used by TransitionParser. A crafted pickle can invoke recfromtxt to read a local file or openmemmap to create or truncate an attacker-selected file when a victim loads an untrusted model. This issue is fixed in version 3.10.4.
Affected Software
Remediation
Recommended actions to resolve this vulnerability, in priority order.
- Upgrade
Upgrade
nltkto a version that resolves this vulnerability.Fixed in 3.10.4
Event History
Frequently Asked Questions
Who is exposed to this issue?
Applications using NLTK versions 3.10.0 through 3.10.4 are exposed when they load an untrusted model through the affected pickle-loading path. Exploitation requires a victim to load a crafted pickle, so systems that do not accept or load untrusted models are not described as affected by this attack scenario.
What can a crafted model do?
A crafted pickle can invoke NumPy's recfromtxt to read a local file, or open_memmap to create or truncate an attacker-selected local file. The stated impact is high confidentiality impact and low integrity impact, with no availability impact.
What should be done if the affected loader is in use?
Update NLTK to version 3.10.4, which contains the fix. Until updating, do not load untrusted models or pickles through the affected NLTK path.