CVE-2025-6638: Transformers ReDoS in MarianTokenizer

CISO Take

Upgrade Hugging Face Transformers to 4.53.0 immediately if your ML stack includes multilingual translation pipelines using MarianMT models. Any internet-facing service that passes untrusted text through MarianTokenizer is vulnerable to CPU exhaustion attacks with no authentication required. The fix is a one-line pip upgrade with no breaking changes.

What is the risk?

Moderate operational risk despite the high CVSS (7.5). EPSS of 0.00032 signals near-zero observed exploitation in the wild, and the vulnerability is limited to availability impact only — no data exfiltration or code execution possible. However, the attack is trivially executable (no auth, no prior access, low complexity) against any exposed translation endpoint, making it a viable DoS vector for motivated adversaries targeting AI-powered services. Risk is elevated for SaaS platforms that expose multilingual NLP APIs publicly.

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
Transformers	pip	—	No patch
161.8K OpenSSF 6.4 8.3K dependents Pushed 4d ago 40% patched ~92d to patch Full package profile →
Transformers	pip	< 4.53.0	`4.53.0`
161.8K OpenSSF 6.4 8.3K dependents Pushed 4d ago 40% patched ~92d to patch Full package profile →

How severe is it?

CVSS 3.1

7.5 / 10

EPSS

0.5%

chance of exploitation in 30 days

Higher than 38% of all CVEs

Source: EPSS v3 — FIRST.org

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?

AV Network

AC Low

PR None

UI None

S Unchanged

C None

I None

A High

What should I do?

5 steps

PATCH (immediate): pip install 'transformers>=4.53.0' — the fix is available and non-breaking.
VERIFY

Run 'pip show transformers | grep Version' across all inference nodes, CI/CD workers, and training environments. Container images built before 4.53.0 release need rebuilding.
WORKAROUND (if patching is delayed): Enforce input length limits upstream (e.g., max 512 chars) before text reaches the tokenizer; validate that input does not contain malformed Unicode or excessive special-character sequences.
DETECT

Monitor CPU spike patterns on translation endpoints; anomalous sustained CPU usage from a single source IP is the primary indicator.
SCOPE CHECK

grep your codebase for 'MarianTokenizer' to identify all usage points.

What does CISA's SSVC say?

Decision Track*

Exploitation poc

Automatable Yes

Technical Impact partial

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

How is it classified?

DoS Supply Chain Framework Inference AML.T0010.001 - AI Software AML.T0029 - Denial of AI Service AML.T0034 - Cost Harvesting AML.T0049 - Exploit Public-Facing Application

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 assessment 8.4 - AI system operation

NIST AI RMF

GOVERN-1.7 - Processes for AI risk from third-party dependencies MANAGE-2.4 - Residual risks are monitored and managed

OWASP LLM Top 10

LLM05 - Supply Chain Vulnerabilities

Frequently Asked Questions

What is CVE-2025-6638?

Upgrade Hugging Face Transformers to 4.53.0 immediately if your ML stack includes multilingual translation pipelines using MarianMT models. Any internet-facing service that passes untrusted text through MarianTokenizer is vulnerable to CPU exhaustion attacks with no authentication required. The fix is a one-line pip upgrade with no breaking changes.

Is CVE-2025-6638 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2025-6638, increasing the risk of exploitation.

How to fix CVE-2025-6638?

1. PATCH (immediate): pip install 'transformers>=4.53.0' — the fix is available and non-breaking. 2. VERIFY: Run 'pip show transformers | grep Version' across all inference nodes, CI/CD workers, and training environments. Container images built before 4.53.0 release need rebuilding. 3. WORKAROUND (if patching is delayed): Enforce input length limits upstream (e.g., max 512 chars) before text reaches the tokenizer; validate that input does not contain malformed Unicode or excessive special-character sequences. 4. DETECT: Monitor CPU spike patterns on translation endpoints; anomalous sustained CPU usage from a single source IP is the primary indicator. 5. SCOPE CHECK: grep your codebase for 'MarianTokenizer' to identify all usage points.

What systems are affected by CVE-2025-6638?

This vulnerability affects the following AI/ML architecture patterns: NLP translation pipelines, model serving, multilingual RAG ingestion, document processing pipelines, batch training pipelines.

What is the CVSS score for CVE-2025-6638?

CVE-2025-6638 has a CVSS v3.1 base score of 7.5 (HIGH). The EPSS exploitation probability is 0.49%.

What is the AI security impact?

Affected AI Architectures

NLP translation pipelinesmodel servingmultilingual RAG ingestiondocument processing pipelinesbatch training pipelines

MITRE ATLAS Techniques

AML.T0010.001 AI Software

AML.T0029 Denial of AI Service

AML.T0034 Cost Harvesting

AML.T0049 Exploit 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-2.4

OWASP LLM Top 10: LLM05

What are the technical details?

Original Advisory

A Regular Expression Denial of Service (ReDoS) vulnerability was discovered in the Hugging Face Transformers library, specifically affecting the MarianTokenizer's `remove_language_code()` method. This vulnerability is present in version 4.52.4 and has been fixed in version 4.53.0. The issue arises from inefficient regex processing, which can be exploited by crafted input strings containing malformed language code patterns, leading to excessive CPU consumption and potential denial of service.

Exploitation Scenario

An adversary identifies a public-facing translation API or document ingestion endpoint powered by HuggingFace Transformers. Using a fuzzing tool or manual crafting, they construct strings with malformed language code patterns — sequences that trigger catastrophic backtracking in the 'remove_language_code()' regex. The adversary sends a modest volume of these payloads concurrently (no flood required — each request saturates a CPU thread). Within seconds, the inference server's CPU reaches 100%, blocking all legitimate requests. For containerized deployments without CPU limits, this can cascade to affect co-located services. No credentials, API keys, or prior knowledge of the model are required.

Weaknesses (CWE)

CWE-1333 Inefficient Regular Expression Complexity Primary CWE-1333 Inefficient Regular Expression Complexity

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.