CVE-2025-1194: transformers: ReDoS in GPT-NeoX Japanese tokenizer

GHSA-fpwr-67px-3qhxMEDIUMPoC AVAILABLECISA: TRACK*
Published April 29, 2025
CISO Take

Upgrade HuggingFace Transformers to 4.50.0 immediately if your stack includes any Japanese NLP workloads. The ReDoS in SubWordJapaneseTokenizer can peg CPU to 100% via a single crafted input, taking down inference services or preprocessing pipelines. If you are not running Japanese language models, your exposure is zero — this is a narrow but real availability risk for those who are.

What is the risk?

Actual risk is low-to-moderate despite CVSS 6.5. EPSS of 0.00078 signals no active exploitation. Attack vector is network but requires user interaction — a downstream user or API consumer must submit the malicious payload to a tokenizer-exposed endpoint. Impact is purely availability (DoS), with no data loss or confidentiality breach. Blast radius is limited to organizations running GPT-NeoX-Japanese models. The fix is available and straightforward (pip upgrade), making residual risk low for patched systems.

What systems are affected?

PackageEcosystemVulnerable RangePatched
Transformerspip—No patch
169.0KOpenSSF 6.59.8K dependentsPushed 3d ago42% patched~91d to patchFull package profile →
Transformerspip< 4.50.04.50.0
169.0KOpenSSF 6.59.8K dependentsPushed 3d ago42% patched~91d to patchFull package profile →

How severe is it?

CVSS 3.1
6.5 / 10
EPSS
0.5%
chance of exploitation in 30 days
Higher than 43% of all CVEs
Exploitation Status
Exploit Available
Exploitation: MEDIUM
Sophistication
Trivial
Exploitation Confidence
medium
○CISA SSVC: Public PoC
○Public PoC indexed (trickest/cve)
Composite signal derived from CISA KEV, VulnCheck KEV, CISA SSVC, EPSS, Metasploit, Exploit-DB, trickest/cve, Nuclei templates, and inthewild.io exploitation reports.

What is the attack surface?

AVACPRUISCIA
AVNetwork
ACLow
PRNone
UIRequired
SUnchanged
CNone
INone
AHigh

What should I do?

5 steps
  1. PATCH

    Upgrade transformers to ≥4.50.0 (pip install --upgrade transformers). This is the only complete fix.

  2. WORKAROUND (if upgrade is blocked): Implement input length caps and character class validation before tokenization; reject inputs exceeding a safe threshold for Japanese text.

  3. DETECTION

    Monitor inference server CPU utilization for sustained spikes correlated with single requests; alert on requests exceeding 2-3x normal tokenization latency.

  4. CONTAINMENT

    If running multi-tenant inference, isolate Japanese tokenizer workloads to dedicated workers with CPU throttling (cgroups/ulimit) to prevent cross-tenant DoS.

  5. VALIDATION

    After patching, confirm version with pip show transformers | grep Version.

What does CISA's SSVC say?

DecisionTrack*
Exploitationpoc
AutomatableNo
Technical Impactpartial

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

How is it classified?

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act
Article 9-Risk management system
ISO 42001
6.1.2-AI risk assessment8.4-AI system operation
NIST AI RMF
GOVERN-1.7-Processes for AI risk and impact trackingMANAGE-2.2-Mechanisms to sustain value of deployed AI systems
OWASP LLM Top 10
LLM04-Model Denial of Service

Frequently Asked Questions

What is CVE-2025-1194?

Upgrade HuggingFace Transformers to 4.50.0 immediately if your stack includes any Japanese NLP workloads. The ReDoS in SubWordJapaneseTokenizer can peg CPU to 100% via a single crafted input, taking down inference services or preprocessing pipelines. If you are not running Japanese language models, your exposure is zero — this is a narrow but real availability risk for those who are.

Is CVE-2025-1194 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2025-1194, increasing the risk of exploitation.

How to fix CVE-2025-1194?

1. PATCH: Upgrade transformers to ≥4.50.0 (pip install --upgrade transformers). This is the only complete fix. 2. WORKAROUND (if upgrade is blocked): Implement input length caps and character class validation before tokenization; reject inputs exceeding a safe threshold for Japanese text. 3. DETECTION: Monitor inference server CPU utilization for sustained spikes correlated with single requests; alert on requests exceeding 2-3x normal tokenization latency. 4. CONTAINMENT: If running multi-tenant inference, isolate Japanese tokenizer workloads to dedicated workers with CPU throttling (cgroups/ulimit) to prevent cross-tenant DoS. 5. VALIDATION: After patching, confirm version with `pip show transformers | grep Version`.

What systems are affected by CVE-2025-1194?

This vulnerability affects the following AI/ML architecture patterns: model serving, training pipelines, NLP processing pipelines, batch inference.

What is the CVSS score for CVE-2025-1194?

CVE-2025-1194 has a CVSS v3.1 base score of 6.5 (MEDIUM). The EPSS exploitation probability is 0.49%.

What is the AI security impact?

Affected AI Architectures

model servingtraining pipelinesNLP processing pipelinesbatch inference

MITRE ATLAS Techniques

AML.T0010.001AI Software
AML.T0029Denial of AI Service
AML.T0034Cost Harvesting
AML.T0049Exploit Public-Facing Application

Compliance Controls Affected

EU AI Act: Article 9
ISO 42001: 6.1.2, 8.4
NIST AI RMF: GOVERN-1.7, MANAGE-2.2
OWASP LLM Top 10: LLM04

What are the technical details?

Original Advisory

A Regular Expression Denial of Service (ReDoS) vulnerability was identified in the huggingface/transformers library, specifically in the file `tokenization_gpt_neox_japanese.py` of the GPT-NeoX-Japanese model. The vulnerability occurs in the SubWordJapaneseTokenizer class, where regular expressions process specially crafted inputs. The issue stems from a regex exhibiting exponential complexity under certain conditions, leading to excessive backtracking. This can result in high CPU usage and potential application downtime, effectively creating a Denial of Service (DoS) scenario. The affected version is v4.48.1 (latest).

Exploitation Scenario

An adversary targeting a Japanese-language sentiment analysis or document processing SaaS API sends a POST request with a specially crafted string designed to trigger catastrophic backtracking in the SubWordJapaneseTokenizer regex engine. No authentication is required if the endpoint is public-facing. The regex processes the input, enters exponential backtracking, and the worker process consumes 100% CPU for an extended period. In a Kubernetes deployment, liveness probes time out and the pod restarts, creating a cycle exploitable to maintain denial of service with low-rate request flooding. The PoC is public via huntr.com, making this accessible to low-sophistication actors targeting Japanese NLP services.

Weaknesses (CWE)

CWE-1333 — Inefficient Regular Expression Complexity: The product uses a regular expression with a worst-case computational complexity that is inefficient and possibly exponential.

  • [Architecture and Design] Use regular expressions that do not support backtracking, e.g. by removing nested quantifiers.
  • [System Configuration] Set backtracking limits in the configuration of the regular expression implementation, such as PHP's pcre.backtrack_limit. Also consider limits on execution time for the process.

Source: MITRE CWE corpus.

CVSS Vector

CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:N/I:N/A:H

Timeline

Published
April 29, 2025
Last Modified
August 4, 2025
First Seen
April 29, 2025

Related Vulnerabilities