Skip to content
TransformersGHSA-4w7r-h757-3r74

Hugging Face Transformers vulnerable to Regular Expression Denial of Service (ReDoS) in the AdamWeightDecay optimizer

Medium5.3CVE-2025-6921 · Published Sep 23, 2025 · updated Sep 10, 2026

GitHub advisory

Affected versions

PackageAffectedFixed in
transformers
PyPI
< 4.53.04.53.0
Details and references

The huggingface/transformers library, versions prior to 4.53.0, is vulnerable to Regular Expression Denial of Service (ReDoS) in the AdamWeightDecay optimizer. The vulnerability arises from the _do_use_weight_decay method, which processes user-controlled regular expressions in the include_in_weight_decay and exclude_from_weight_decay lists. Malicious regular expressions can cause catastrophic backtracking during the re.search call, leading to 100% CPU utilization and a denial of service. This issue can be exploited by attackers who can control the patterns in these lists, potentially causing the machine learning task to hang and rendering services unresponsive.

CVSS 3.0
CVSS:3.0/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:L
Severity from
GitHub (reviewed advisory)
Weakness
CWE-400
Also known as
CVE-2025-6921, PYSEC-2026-1980

More Transformers advisories

All Transformers
DateAdvisory
Sep 142025Hugging Face Transformers library has Regular Expression Denial of Service
CVE-2025-6051Medium5.3fixed in 4.53.0
Sep 122025Hugging Face Transformers is vulnerable to ReDoS through its MarianTokenizer
CVE-2025-6638Medium5.3fixed in 4.53.0
Aug 62025Hugging Face Transformers Regular Expression Denial of Service (ReDoS) vulnerability
CVE-2025-5197Medium5.3fixed in 4.53.0
Jul 112025Transformers is vulnerable to ReDoS attack through its DonutProcessor class
CVE-2025-3933Medium5.3fixed in 4.52.1
Jul 72025Transformers vulnerable to ReDoS attack through its SETTING_RE variable
CVE-2025-3262Medium5.3fixed in 4.51.0
Jul 72025Transformers vulnerable to ReDoS attack through its get_imports() function
CVE-2025-3264Medium5.3fixed in 4.51.0

Critical advisories by email

Wednesdays: the week’s critical and high advisories in the AI and data stack, with the fixed versions. Only in weeks that have some.

Double opt-in. Unsubscribe any time.