CVE-2025-62608: mlx: security flaw enables exploitation

GHSA-w6vg-jg77-2qg6CRITICALPoC AVAILABLECISA: TRACK*

Published November 21, 2025

CISO Take

MLX, Apple's ML framework for Apple Silicon, has a critical heap buffer overflow triggered by malicious NumPy .npy files — no authentication or user interaction required. Any team using MLX to load externally-sourced datasets or model weights on Apple hardware is exposed to heap memory disclosure and potential RCE. Patch to 0.29.4 immediately and audit pipelines that ingest .npy files from untrusted sources.

What is the risk?

CVSS 9.1 Critical with network-reachable, zero-privilege, zero-interaction attack vector makes this technically severe. Practical risk is moderated by low EPSS (0.00095) indicating limited current exploitation activity, and the specific hardware constraint (Apple Silicon). However, MLX adoption is growing rapidly among ML engineers and on-device AI developers — orgs with Apple Silicon MacBooks in ML pipelines have meaningful exposure, especially in dev/research environments where data provenance controls are weak.

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
MLX	pip	—	No patch
29.0K429 dependentsPushed 6d ago40% patched~11d to patchFull package profile →
MLX	pip	<= 0.29.3	`0.29.4`
29.0K429 dependentsPushed 6d ago40% patched~11d to patchFull package profile →

How severe is it?

CVSS 3.1

9.1 / 10

EPSS

0.6%

chance of exploitation in 30 days

Higher than 47% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

Exploit Available

Exploitation: MEDIUM

Sophistication

Moderate

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?

AVNetwork

ACLow

PRNone

UINone

SUnchanged

CHigh

INone

AHigh

What should I do?

6 steps

PATCH
Upgrade mlx to >= 0.29.4 via pip immediately — this is a one-line fix with no breaking changes.
INVENTORY
Identify all systems running MLX (pip list | grep mlx on Apple Silicon endpoints and CI runners).
INPUT VALIDATION
Implement file origin controls — reject .npy files not sourced from trusted, internal artifact registries; add hash verification for dataset files.
ISOLATION
Run MLX data loading in sandboxed processes where possible; restrict access to sensitive memory (credentials, keys) in processes that handle external .npy files.
DETECTION
Monitor for abnormal crashes in mlx processes, unexpected heap access violations in ML workloads, and unusual .npy file sources in pipeline logs.
CI/CD: Add MLX version pinning checks to CI pipelines to prevent regression to vulnerable versions.

What does CISA's SSVC say?

DecisionTrack*

Exploitationpoc

AutomatableYes

Technical Impactpartial

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

How is it classified?

Supply ChainCode ExecutionData ExtractionFrameworkTraining DataAML.T0002.000- Datasets AML.T0010.001- AI Software AML.T0010.002- Data AML.T0011.000- Unsafe AI Artifacts AML.T0025- Exfiltration via Cyber Means AML.T0049- Exploit Public-Facing Application

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Art. 15-Accuracy, robustness and cybersecurity for high-risk AI systemsArt. 9-Risk management systemArticle 15-Accuracy, robustness and cybersecurity

ISO 42001

A.6.1.2-Information security in AI system developmentA.6.2-AI system life cycle — supplier relationshipsA.8.4-AI supply chain management

NIST AI RMF

GV-6.1-Policies for AI risk identification and management of third-party dependenciesMANAGE 2.2-Mechanisms to respond to and recover from AI risksMS-2.5-Practices and personnel for AI risk management

OWASP LLM Top 10

LLM03-Training Data PoisoningLLM05:2025-Supply Chain VulnerabilitiesLLM06:2025-Sensitive Information Disclosure

Frequently Asked Questions

What is CVE-2025-62608?

Is CVE-2025-62608 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2025-62608, increasing the risk of exploitation.

How to fix CVE-2025-62608?

1. PATCH: Upgrade mlx to >= 0.29.4 via pip immediately — this is a one-line fix with no breaking changes. 2. INVENTORY: Identify all systems running MLX (pip list | grep mlx on Apple Silicon endpoints and CI runners). 3. INPUT VALIDATION: Implement file origin controls — reject .npy files not sourced from trusted, internal artifact registries; add hash verification for dataset files. 4. ISOLATION: Run MLX data loading in sandboxed processes where possible; restrict access to sensitive memory (credentials, keys) in processes that handle external .npy files. 5. DETECTION: Monitor for abnormal crashes in mlx processes, unexpected heap access violations in ML workloads, and unusual .npy file sources in pipeline logs. 6. CI/CD: Add MLX version pinning checks to CI pipelines to prevent regression to vulnerable versions.

What systems are affected by CVE-2025-62608?

This vulnerability affects the following AI/ML architecture patterns: training pipelines, data preprocessing pipelines, model fine-tuning workflows, on-device inference, MLOps CI/CD pipelines.

What is the CVSS score for CVE-2025-62608?

CVE-2025-62608 has a CVSS v3.1 base score of 9.1 (CRITICAL). The EPSS exploitation probability is 0.56%.

What is the AI security impact?

Affected AI Architectures

training pipelinesdata preprocessing pipelinesmodel fine-tuning workflowson-device inferenceMLOps CI/CD pipelines

MITRE ATLAS Techniques

AML.T0002.000Datasets

AML.T0010.001AI Software

AML.T0010.002Data

AML.T0011.000Unsafe AI Artifacts

AML.T0025Exfiltration via Cyber Means

AML.T0049Exploit Public-Facing Application

Compliance Controls Affected

EU AI Act: Art. 15, Art. 9, Article 15

ISO 42001: A.6.1.2, A.6.2, A.8.4

NIST AI RMF: GV-6.1, MANAGE 2.2, MS-2.5

OWASP LLM Top 10: LLM03, LLM05:2025, LLM06:2025

What are the technical details?

Original Advisory

MLX is an array framework for machine learning on Apple silicon. Prior to version 0.29.4, there is a heap buffer overflow in mlx::core::load() when parsing malicious NumPy .npy files. Attacker-controlled file causes 13-byte out-of-bounds read, leading to crash or information disclosure. This issue has been patched in version 0.29.4.

Exploitation Scenario

Adversary targets an ML engineering team's data ingestion pipeline. They publish a poisoned NumPy dataset to a public repository (Hugging Face, Kaggle, or GitHub) or compromise an internal data storage bucket. A data scientist or automated training pipeline calls mlx.load() on the malicious .npy file on their Apple Silicon MacBook or M-series Mac mini training node. The crafted file triggers the heap buffer overflow in mlx::core::load(), causing a 13-byte out-of-bounds read that leaks adjacent heap memory — potentially exposing API keys, model weights, or other secrets loaded earlier in the process. In a CI/CD context with automated training jobs, this could be fully unattended exploitation.

Weaknesses (CWE)

CWE-122Heap-based Buffer OverflowPrimary CWE-122Heap-based Buffer Overflow

CWE-122 — Heap-based Buffer Overflow: The product writes past the end of a buffer that was allocated on the heap.

Pre-design: Use a language or compiler that performs automatic bounds checking.
[Architecture and Design] Use an abstraction library to abstract away risky APIs. Not a complete solution.

Source: MITRE CWE corpus.