Transformers save_pretrained path traversal allows arbitrary file writes through chat template names
CVE-2026-9856 · GHSA-xrqw-3rrv-vx5w · published 2026-08-02 · updated 2026-09-01
- Severity
- High
- CVSS score
- 7.1
- Chance of exploitation (EPSS)
- 0.46%
- Fixed in
- 5.10.0
- Exploitation No known exploitation
- Automatable No
- Technical impact Partial
- CISA SSVC
A vulnerability in huggingface/transformers versions < 5.10.0 allows an attacker to perform arbitrary file writes via path traversal. The issue resides in the savepretrained() methods of PreTrainedTokenizerBase and ProcessorMixin, where keys from the chattemplate 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.
Upgrade transformers to 5.10.0 or later.
| Package | Registry | Affected | Fixed in |
|---|---|---|---|
| transformers | PyPI | < 5.10.0 | 5.10.0 |
- CriticalMCP Path Traversal Allows Full-Filesystem Read and JWT Secret Exfiltration Leading to Account Takeoverlangflow, langflow-base
- Critical@zereight/mcp-gitlab: Unauthenticated arbitrary file read via `upload_markdown` enables PAT exfiltration and full account takeover@zereight/mcp-gitlab
- CriticalFlowise RCE via SQL Database Chain Nodeflowise, flowise-components
- CriticalLangroid: SQLChatAgent dangerous-function blocklist can be bypassed with quoted or schema-qualified pg_read_file callslangroid
- Criticalfast-mcp-telegram: Bearer token path traversal bypasses reserved Telegram session protectionfast-mcp-telegram
- CriticalMCP Toolbox HTTP tool path parameters allow traversal to unintended endpointsgithub.com/googleapis/mcp-toolbox
- Criticaltransformers has a Deserialization of Untrusted Data vulnerabilitytransformers
- HighHugging Face Transformers downloads custom generation code before trust consenttransformers
- Highhuggingface/transformers: Arbitrary Code Execution During Model Initialization in the LightGlue Model Loading Pathtransformers
- HighHuggingFace transformers vulnerable to remote code executiontransformers
- HighDeserialization of Untrusted Data in Hugging Face Transformerstransformers
- HighDeserialization of Untrusted Data in Hugging Face Transformerstransformers
The impact, patch and workaround text comes from the advisory, shortened (GitHub Advisory Database, CC BY 4.0). Proof-of-concept and reproduction sections and code are left out; the original advisory has them.