CVE-2023-6730: Transformers RCE — HIGH

CISO Take

Any team loading HuggingFace models via the transformers library before 4.36.0 is exposed to remote code execution — triggered simply by loading a malicious model file. Patch to 4.36.0 immediately and audit all model-loading pipelines for untrusted sources. This is a supply-chain RCE vector that bypasses application-layer controls entirely.

What is the risk?

HIGH operational risk for organizations with active ML pipelines. CVSS 8.8 with network vector, low complexity, and low privilege requirements makes exploitation straightforward once an attacker can position a malicious model file in the loading path. EPSS is currently low (0.16%) suggesting limited automated scanning, but the technique is well-understood by the offensive community and requires no AI expertise. Exposure scales with how many teams load models from shared registries, CI/CD pipelines, or external sources.

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	< 4.36.0	`4.36.0`
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

8.8 / 10

EPSS

0.9%

chance of exploitation in 30 days

Higher than 56% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

Exploit Available

Exploitation: MEDIUM

Sophistication

Moderate

Exploitation Confidence

medium

○ 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 Low

UI None

S Unchanged

C High

I High

A High

What should I do?

6 steps

PATCH

Upgrade transformers to >= 4.36.0 immediately. Verify via pip show transformers.
AUDIT

Inventory all code calling from_pretrained() or equivalent — flag any loading from non-verified sources.
ARTIFACT TRUST

Enforce model hash pinning (SHA256) for all production model pulls; reject unsigned or unverified model files.
FORMAT

Migrate to safetensors format where possible — it eliminates pickle-based deserialization entirely.
ISOLATION

Run model loading in sandboxed environments (containers with no network egress, minimal filesystem permissions).
DETECT

Alert on unexpected outbound connections from model-serving processes post-load.

How is it classified?

Supply Chain Code Execution Framework Model AML.T0010.001 - AI Software AML.T0010.003 - Model AML.T0011.000 - Unsafe AI Artifacts AML.T0018.002 - Embed Malware AML.T0058 - Publish Poisoned Models

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 17 - Quality management system — third-party components Article 9 - Risk management system

ISO 42001

6.1.2 - AI risk assessment 8.4 - AI system supply chain

NIST AI RMF

GOVERN 1.2 - Accountability for AI risk MANAGE 2.2 - Treatment of identified AI risks

OWASP LLM Top 10

LLM03:2025 - Supply Chain Vulnerabilities

Frequently Asked Questions

What is CVE-2023-6730?

Any team loading HuggingFace models via the transformers library before 4.36.0 is exposed to remote code execution — triggered simply by loading a malicious model file. Patch to 4.36.0 immediately and audit all model-loading pipelines for untrusted sources. This is a supply-chain RCE vector that bypasses application-layer controls entirely.

Is CVE-2023-6730 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2023-6730, increasing the risk of exploitation.

How to fix CVE-2023-6730?

1. PATCH: Upgrade transformers to >= 4.36.0 immediately. Verify via `pip show transformers`. 2. AUDIT: Inventory all code calling from_pretrained() or equivalent — flag any loading from non-verified sources. 3. ARTIFACT TRUST: Enforce model hash pinning (SHA256) for all production model pulls; reject unsigned or unverified model files. 4. FORMAT: Migrate to safetensors format where possible — it eliminates pickle-based deserialization entirely. 5. ISOLATION: Run model loading in sandboxed environments (containers with no network egress, minimal filesystem permissions). 6. DETECT: Alert on unexpected outbound connections from model-serving processes post-load.

What systems are affected by CVE-2023-6730?

This vulnerability affects the following AI/ML architecture patterns: training pipelines, model serving, fine-tuning pipelines, RAG pipelines, agent frameworks, MLOps CI/CD pipelines.

What is the CVSS score for CVE-2023-6730?

CVE-2023-6730 has a CVSS v3.1 base score of 8.8 (HIGH). The EPSS exploitation probability is 0.92%.

What is the AI security impact?

Affected AI Architectures

training pipelinesmodel servingfine-tuning pipelinesRAG pipelinesagent frameworksMLOps CI/CD pipelines

MITRE ATLAS Techniques

AML.T0010.001 AI Software

AML.T0010.003 Model

AML.T0011.000 Unsafe AI Artifacts

AML.T0018.002 Embed Malware

AML.T0058 Publish Poisoned Models

Compliance Controls Affected

EU AI Act: Article 17, Article 9

ISO 42001: 6.1.2, 8.4

NIST AI RMF: GOVERN 1.2, MANAGE 2.2

OWASP LLM Top 10: LLM03:2025

What are the technical details?

Original Advisory

Deserialization of Untrusted Data in GitHub repository huggingface/transformers prior to 4.36.

Exploitation Scenario

Adversary creates a malicious HuggingFace model repository with a crafted pickle payload embedded in the model weights file (e.g., pytorch_model.bin). The payload establishes a reverse shell or exfiltrates credentials upon deserialization. The attacker promotes the repository via SEO, GitHub stars, or direct targeting of a victim org's model sourcing workflow. When a developer or automated MLOps pipeline runs `AutoModel.from_pretrained('attacker/malicious-model')`, the payload executes in the context of the training or inference server — often with cloud credentials or internal network access. No interaction beyond the model pull is required.

Weaknesses (CWE)

CWE-502 Deserialization of Untrusted Data Primary CWE-502 Deserialization of Untrusted Data

CWE-502 — Deserialization of Untrusted Data: The product deserializes untrusted data without sufficiently ensuring that the resulting data will be valid.

[Architecture and Design, Implementation] If available, use the signing/sealing features of the programming language to assure that deserialized data has not been tainted. For example, a hash-based message authentication code (HMAC) could be used to ensure that data has not been modified.
[Implementation] When deserializing data, populate a new object rather than just deserializing. The result is that the data flows through safe input validation and that the functions are safe.

Source: MITRE CWE corpus.