CVE-2026-80047
HighCVSS 7.8Exploitation Probability (EPSS)
Low risk1th percentile - higher than 1% of all known CVEs
Summary
A vulnerability in Hugging Face Transformers (versions 4.57.0 to 5.16.1) allows remote Python files to be written to local disk without user consent when using GenerativePreTrainedModel.load_custom_generate(). The function fetches and caches a remote module file before performing the required trust_remote_code consent check, inverting the security model enforced by other code-loading paths. As a result, attacker-controlled Python code from custom_generate/generate.py is copied into the user's ~/.cache/huggingface/modules directory even if the user declines the trust prompt.
Risk Assessment
Persistent, unauthorized files may remain on disk, and cached attacker code may later be executed during trusted model loads, leading to remote code execution.
Recommendation
Update Hugging Face Transformers to a version later than 5.16.1, and until patched, avoid using load_custom_generate() and clear the ~/.cache/huggingface/modules directory.
Other vulnerabilities in Hugging Face Transformers
See all- CVE-2026-75104Medium
Hugging Face Transformers fails to validate shard filenames in checkpoint index files, allowing attackers to read arbitrary files outside the model directory. Malicious index files with parent-directory references or absolute paths can lead to file disclosure and filesystem reconnaissance.
- CVE-2026-9856High
Hugging Face Transformers up to version 5.8.0.dev0 has an arbitrary file write vulnerability via path traversal in save_pretrained(). Keys from chat_template are used as filenames without validation.
- CVE-2025-14930High
Deserialization of untrusted data in Hugging Face Transformers for the GLM4 model when parsing weights. Lack of validation of user-supplied data can lead to deserialization of untrusted data and remote code execution.
- CVE-2025-14929High
A vulnerability in Hugging Face Transformers' X-CLIP checkpoint conversion involves deserialization of untrusted data without proper validation. Exploitation requires user interaction (opening a malicious file or visiting a malicious page). It can lead to remote code execution in the context of the current process.
- CVE-2025-14928High
Code injection in the convert_config function of Hugging Face Transformers for the HuBERT model. Lack of validation of a user-supplied string before using it to execute Python code allows remote code execution.
- CVE-2025-14927High
Code injection in the convert_config function of Hugging Face Transformers for the SEW-D model. Lack of validation of a user-supplied string before using it to execute Python code allows remote code execution.
- CVE-2025-14926High
Code injection in the convert_config function of Hugging Face Transformers for the SEW model. Lack of validation of a user-supplied string before using it to execute Python code allows remote code execution.
- CVE-2025-14924High
Deserialization of untrusted data in Hugging Face Transformers for the megatron_gpt2 model when parsing checkpoints. Lack of validation of user-supplied data can lead to deserialization of untrusted data and remote code execution.
- CVE-2025-14921High
Deserialization of untrusted data in Hugging Face Transformers for the Transformer-XL model when parsing model files. Lack of validation of user-supplied data can lead to deserialization of untrusted data and remote code execution.
- CVE-2025-14920High
A vulnerability in Hugging Face Transformers' Perceiver model parsing involves deserialization of untrusted data without proper validation. Exploitation requires user interaction (opening a malicious file or visiting a malicious page). It can lead to remote code execution in the context of the current user.
Original NVD description (English source)
A vulnerability in Hugging Face Transformers (versions 4.57.0 to 5.16.1) allows remote Python files to be written to local disk without user consent when using GenerativePreTrainedModel.load_custom_generate(). The function fetches and caches a remote module file before performing the required trust_remote_code consent check, inverting the security model enforced by other code-loading paths (such as AutoConfig, AutoModel, and AutoTokenizer). As a result, attacker‑controlled Python code from custom_generate/generate.py is copied into the user’s ~/.cache/huggingface/modules directory even if the user declines the trust prompt. Although execution is correctly gated, the file write is not reversible and can persist across sessions. This can lead to persistent, unauthorized files on disk and stale cache collisions where cached attacker code may later be executed during trusted model loads. The issue stems from an unconditional file write in dynamic_module_utils.py prior to any trust verification.
Vulnerability data from NVD (NIST) · CISA KEV · EPSS