CVE-2025-5197: Hugging Face Transformers Regular Expression Denial of Service (ReDoS) vulnerability
The Hugging Face Transformers function `convert_tf_weight_name_to_pt_weight_name()` uses the regex `/[^/]*___([^/]*)/`, which backtracks catastrophically on crafted names. A long weight name can pin a CPU during TensorFlow-to-PyTorch conversion. Versions before 4.53.0 are affected, and 4.53.0 fixes it. Impact is limited to availability.
Published Aug 6, 2025 · Analysis Nov 1, 2026
What an attacker can do
The function converts TensorFlow weight names to PyTorch names in modeling_tf_pytorch_utils.py. It applies the regular expression /[^/]*___([^/]*)/ with re.sub. On a long name with many ___ separators the regex engine backtracks, so CPU time grows far faster than input length. One crafted name can hold a core for a very long time.
The attacker must be able to supply the weight names, for example through a TensorFlow checkpoint or model file that a service converts, or an API that exposes model conversion. No authentication or user action is needed in that setup (CVSS PR:N/UI:N). The result is service disruption and resource exhaustion, scored as low availability impact. The sources describe no data exposure or code execution. Services that never convert TensorFlow checkpoints do not reach the function.
Who is affected
The transformers pip package before 4.53.0, in code paths that load TensorFlow weights into PyTorch models (for example from_pretrained(..., from_tf=True)) or call the conversion helper directly. The advisory says versions up to 4.51.3 are affected. Check:
pip show transformersSearch your code for from_tf=True and convert_tf_weight_name_to_pt_weight_name, and note whether untrusted checkpoints reach them.
transformers
- Ecosystem
- PyPI
- Affected
- < 4.53.0
- Fixed in
- 4.53.0
- Check
pip show transformers
How to fix it
Upgrade to transformers 4.53.0:
pip install --upgrade "transformers>=4.53.0"The fix adds a guard before the re.sub call, labelled "ReDOS check". It raises a ValueError with the message "TF variable name is too long or contains too many ___ separators" for names that are over 2048 characters and contain ___, or that contain ___ more than 10 times. The sources list no workaround. Until you upgrade, avoid converting TensorFlow checkpoints from untrusted sources.
How to tell if you were hit
Look for conversion jobs or workers that sit at 100% CPU for a long time during TensorFlow-to-PyTorch loading, and for timeouts in services that expose model conversion. In patched versions, a rejected name logs the ValueError text "TF variable name is too long or contains too many ___ separators". Inspect the checkpoint in question for variable names with very long strings or many triple underscores.
Original advisory
From the GitHub Advisory Database (CC BY 4.0).
A Regular Expression Denial of Service (ReDoS) vulnerability exists in the Hugging Face Transformers library, specifically in the convert_tf_weight_name_to_pt_weight_name() function. This function, responsible for converting TensorFlow weight names to PyTorch format, uses a regex pattern /[^/]*___([^/]*)/ that can be exploited to cause excessive CPU consumption through crafted input strings due to catastrophic backtracking. The vulnerability affects versions up to 4.51.3 and is fixed in version 4.53.0. This issue can lead to service disruption, resource exhaustion, and potential API service vulnerabilities, impacting model conversion processes between TensorFlow and PyTorch formats.
Questions
- Which Transformers version fixes CVE-2025-5197?
4.53.0. The advisory text says versions up to 4.51.3 are affected, and the package data lists everything before 4.53.0.
- Am I affected if I only use PyTorch weights?
The flawed function converts TensorFlow weight names, so code that only loads PyTorch or safetensors weights does not normally call it. Check for
from_tf=Truein your code.- Can this lead to code execution?
No. The sources describe only excessive CPU use from regex backtracking, with low availability impact in the CVSS vector.
References
How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. The analysis is written by an AI model from the sources listed above and checked by automated rules before it is published. Check versions against the vendor's advisory before you act. Methodology
More Transformers vulnerabilities
All Transformers entriesCVE-2026-80047
Hugging Face Transformers downloads custom generation code before trust consent
high7.8No fixSep 1, 2026CVE-2026-9856
Transformers save_pretrained path traversal allows arbitrary file writes through chat template names
high7.1Aug 2, 2026CVE-2026-5241
huggingface/transformers: Arbitrary Code Execution During Model Initialization in the LightGlue Model Loading Path
high8.0Jun 3, 2026CVE-2026-4372
HuggingFace transformers vulnerable to remote code execution
high7.8May 26, 2026CVE-2026-1839
HuggingFace Transformers allows for arbitrary code execution in the `Trainer` class
medium6.5Apr 7, 2026CVE-2025-6921
Hugging Face Transformers vulnerable to Regular Expression Denial of Service (ReDoS) in the AdamWeightDecay optimizer
medium5.3Sep 23, 2025
