PYSEC-2026-3929
Vulnerability from pysec - Published: 2026-09-10 09:44 - Updated: 2026-09-10 11:02A vulnerability in huggingface/transformers versions < 5.10.0 allows an attacker to perform arbitrary file writes via path traversal. The issue resides in the save_pretrained() methods of PreTrainedTokenizerBase and ProcessorMixin, where keys from the chat_template dictionary are used directly as filenames without proper validation. An attacker can exploit this by publishing a malicious Hugging Face Hub repository with a crafted tokenizer_config.json file. When a victim downloads and saves the tokenizer or processor, the attacker-controlled keys can escape the intended save directory, enabling arbitrary file writes with attacker-controlled content. This vulnerability affects multiple processors inheriting from ProcessorMixin, including Idefics, Florence, Gemma, Phi, and Qwen-VL.
| Name | purl | transformers | pkg:pypi/transformers |
|---|
{
"affected": [
{
"package": {
"ecosystem": "PyPI",
"name": "transformers",
"purl": "pkg:pypi/transformers"
},
"ranges": [
{
"events": [
{
"introduced": "0"
},
{
"fixed": "5.10.0"
}
],
"type": "ECOSYSTEM"
}
],
"versions": [
"0.1",
"2.0.0",
"2.1.0",
"2.1.1",
"2.10.0",
"2.11.0",
"2.2.0",
"2.2.1",
"2.2.2",
"2.3.0",
"2.4.0",
"2.4.1",
"2.5.0",
"2.5.1",
"2.6.0",
"2.7.0",
"2.8.0",
"2.9.0",
"2.9.1",
"3.0.0",
"3.0.1",
"3.0.2",
"3.1.0",
"3.2.0",
"3.3.0",
"3.3.1",
"3.4.0",
"3.5.0",
"3.5.1",
"4.0.0",
"4.0.0rc1",
"4.0.1",
"4.1.0",
"4.1.1",
"4.10.0",
"4.10.1",
"4.10.2",
"4.10.3",
"4.11.0",
"4.11.1",
"4.11.2",
"4.11.3",
"4.12.0",
"4.12.1",
"4.12.2",
"4.12.3",
"4.12.4",
"4.12.5",
"4.13.0",
"4.14.0",
"4.14.1",
"4.15.0",
"4.16.0",
"4.16.1",
"4.16.2",
"4.17.0",
"4.18.0",
"4.19.0",
"4.19.1",
"4.19.2",
"4.19.3",
"4.19.4",
"4.2.0",
"4.2.1",
"4.2.2",
"4.20.0",
"4.20.1",
"4.21.0",
"4.21.1",
"4.21.2",
"4.21.3",
"4.22.0",
"4.22.1",
"4.22.2",
"4.23.0",
"4.23.1",
"4.24.0",
"4.25.0",
"4.25.1",
"4.26.0",
"4.26.1",
"4.27.0",
"4.27.1",
"4.27.2",
"4.27.3",
"4.27.4",
"4.28.0",
"4.28.1",
"4.29.0",
"4.29.1",
"4.29.2",
"4.3.0",
"4.3.0rc1",
"4.3.1",
"4.3.2",
"4.3.3",
"4.30.0",
"4.30.1",
"4.30.2",
"4.31.0",
"4.32.0",
"4.32.1",
"4.33.0",
"4.33.1",
"4.33.2",
"4.33.3",
"4.34.0",
"4.34.1",
"4.35.0",
"4.35.1",
"4.35.2",
"4.36.0",
"4.36.1",
"4.36.2",
"4.37.0",
"4.37.1",
"4.37.2",
"4.38.0",
"4.38.1",
"4.38.2",
"4.39.0",
"4.39.1",
"4.39.2",
"4.39.3",
"4.4.0",
"4.4.1",
"4.4.2",
"4.40.0",
"4.40.1",
"4.40.2",
"4.41.0",
"4.41.1",
"4.41.2",
"4.42.0",
"4.42.1",
"4.42.2",
"4.42.3",
"4.42.4",
"4.43.0",
"4.43.1",
"4.43.2",
"4.43.3",
"4.43.4",
"4.44.0",
"4.44.1",
"4.44.2",
"4.45.0",
"4.45.1",
"4.45.2",
"4.46.0",
"4.46.1",
"4.46.2",
"4.46.3",
"4.47.0",
"4.47.1",
"4.48.0",
"4.48.1",
"4.48.2",
"4.48.3",
"4.49.0",
"4.5.0",
"4.5.1",
"4.50.0",
"4.50.1",
"4.50.2",
"4.50.3",
"4.51.0",
"4.51.1",
"4.51.2",
"4.51.3",
"4.52.0",
"4.52.1",
"4.52.2",
"4.52.3",
"4.52.4",
"4.53.0",
"4.53.1",
"4.53.2",
"4.53.3",
"4.54.0",
"4.54.1",
"4.55.0",
"4.55.1",
"4.55.2",
"4.55.3",
"4.55.4",
"4.56.0",
"4.56.1",
"4.56.2",
"4.57.0",
"4.57.1",
"4.57.2",
"4.57.3",
"4.57.4",
"4.57.5",
"4.57.6",
"4.6.0",
"4.6.1",
"4.7.0",
"4.8.0",
"4.8.1",
"4.8.2",
"4.9.0",
"4.9.1",
"4.9.2",
"5.0.0",
"5.0.0rc0",
"5.0.0rc1",
"5.0.0rc2",
"5.0.0rc3",
"5.1.0",
"5.2.0",
"5.3.0",
"5.4.0",
"5.5.0",
"5.5.1",
"5.5.2",
"5.5.3",
"5.5.4",
"5.6.0",
"5.6.1",
"5.6.2",
"5.7.0",
"5.8.0",
"5.8.1",
"5.9.0"
]
}
],
"aliases": [
"CVE-2026-9856",
"GHSA-xrqw-3rrv-vx5w"
],
"details": "A vulnerability in huggingface/transformers versions \u003c 5.10.0 allows an attacker to perform arbitrary file writes via path traversal. The issue resides in the `save_pretrained()` methods of `PreTrainedTokenizerBase` and `ProcessorMixin`, where keys from the `chat_template` dictionary are used directly as filenames without proper validation. An attacker can exploit this by publishing a malicious Hugging Face Hub repository with a crafted `tokenizer_config.json` file. When a victim downloads and saves the tokenizer or processor, the attacker-controlled keys can escape the intended save directory, enabling arbitrary file writes with attacker-controlled content. This vulnerability affects multiple processors inheriting from `ProcessorMixin`, including Idefics, Florence, Gemma, Phi, and Qwen-VL.",
"id": "PYSEC-2026-3929",
"modified": "2026-09-10T11:02:34.656018Z",
"published": "2026-09-10T09:44:51.053247Z",
"references": [
{
"type": "ADVISORY",
"url": "https://nvd.nist.gov/vuln/detail/CVE-2026-9856"
},
{
"type": "WEB",
"url": "https://github.com/huggingface/transformers/pull/46191"
},
{
"type": "WEB",
"url": "https://github.com/huggingface/transformers/commit/eaaaf8494dd5386634ae37d1d122212fdc315be5"
},
{
"type": "PACKAGE",
"url": "https://github.com/huggingface/transformers"
},
{
"type": "WEB",
"url": "https://huntr.com/bounties/362824d5-fe18-40e8-a6cf-62277f97a170"
},
{
"type": "PACKAGE",
"url": "https://pypi.org/project/transformers"
},
{
"type": "ADVISORY",
"url": "https://github.com/advisories/GHSA-xrqw-3rrv-vx5w"
}
],
"severity": [
{
"score": "CVSS:3.0/AV:N/AC:L/PR:N/UI:R/S:U/C:N/I:H/A:L",
"type": "CVSS_V3"
}
],
"summary": "Transformers save_pretrained path traversal allows arbitrary file writes through chat template names"
}
Sightings
| Author | Source | Type | Date | Other |
|---|
Nomenclature
- Seen: The vulnerability was mentioned, discussed, or observed by the user.
- Confirmed: The vulnerability has been validated from an analyst's perspective.
- Published Proof of Concept: A public proof of concept is available for this vulnerability.
- Exploited: The vulnerability was observed as exploited by the user who reported the sighting.
- Patched: The vulnerability was observed as successfully patched by the user who reported the sighting.
- Not exploited: The vulnerability was not observed as exploited by the user who reported the sighting.
- Not confirmed: The user expressed doubt about the validity of the vulnerability.
- Not patched: The vulnerability was not observed as successfully patched by the user who reported the sighting.
The approach is described in our paper Mapping CVEs to MITRE ATT&CK Techniques: A Curated Gold-Set Classifier and the Limits of LLM-Assisted Label Expansion.