Cyber Resilience

CVE-2025-24357

RCE in Vllm ≤ 0.7.0

Published
27 January 2025
Modified
27 June 2025
Patch / advisory
CVSS Score v3.1 7.5
Click a component to see what it means
Raw vectorCVSS:3.1/AV:N/AC:H/PR:N/UI:R/S:U/C:H/I:H/A:H
EPSS Score 0.0071 50th percentile
Risk Priority 57 floored blend · peak EPSS

Summary

CVE-2025-24357 is a high-severity Deserialization of Untrusted Data (CWE-502) vulnerability in Vllm Vllm. Its CVSS base score is 7.5 (High).

Operationally, exploitation aligns with the MITRE ATT&CK technique Exploit Public-Facing Application (T1190); ranked in the top 50% of CVEs by exploit likelihood; it is not currently listed in the CISA KEV catalog.

This vulnerability is AI-related — categorised as NLP and Transformers; in the Supply Chain and Deployment risk domain.

The strongest mitigations our analysis identified map to SA-11 (Developer Testing and Evaluation) and SI-10 (Information Input Validation) — see the control section below for these in your framework.

Deeper analysis AI-assisted summary

Synthesised by an AI model from the NVD description and linked references — a reading aid, not an authoritative source.

vLLM is a library for LLM inference and serving that is affected by CVE-2025-24357 in the hf_model_weights_iterator function within vllm/model_executor/weight_utils.py. This component downloads model checkpoints from Hugging Face and loads them using torch.load with the weights_only parameter defaulting to False, which permits execution of arbitrary code when unpickling malicious data. The issue is a deserialization flaw tracked as CWE-502 and was corrected in release v0.7.0.

An attacker can publish or substitute a malicious checkpoint on Hugging Face that a user subsequently loads through vLLM. Because the load occurs over the network with no requirement for local privileges beyond fetching the model, successful exploitation yields arbitrary code execution on the target system and can compromise confidentiality, integrity, and availability.

Project advisories and the referenced commit indicate that upgrading to v0.7.0 resolves the vulnerability by ensuring safe loading behavior for model weights. The associated GitHub security advisory and pull request document the change and link to PyTorch documentation on torch.load for further context on the underlying parameter.

The affected software is used in AI/ML inference pipelines, making the unsafe default particularly relevant to practitioners handling externally sourced models.

OWASP Top 10 for Web (2025)

EU & UK References

Vulnerability Data

vLLM is a library for LLM inference and serving. vllm/model_executor/weight_utils.py implements hf_model_weights_iterator to load the model checkpoint, which is downloaded from huggingface. It uses the torch.load function and the weights_only parameter defaults to False. When torch.load loads malicious pickle data,…

more

it will execute arbitrary code during unpickling. This vulnerability is fixed in v0.7.0.

CWE(s)

AI Security AnalysisAI

AI Category
NLP and Transformers
Risk Domain
Supply Chain and Deployment
OWASP Top 10 for LLMs 2025
None mapped
Classification Reason
Matched keywords: huggingface, llm, vllm

Related Threats

MITRE ATT&CK Enterprise Techniques

T1190 Exploit Public-Facing Application Initial Access
Adversaries may attempt to exploit a weakness in an Internet-facing host or system to initially access a network.
T1203 Exploitation for Client Execution Execution
Adversaries may exploit software vulnerabilities in client applications to execute code.
T1210 Exploitation of Remote Services Lateral Movement
Adversaries may exploit remote services to gain unauthorized access to internal systems once inside of a network.
Derived from this CVE’s CWE(s) via the direct CWE→ATT&CK cross-walk.

CVEs Like This One

CVE-2025-30165Same product: Vllm Vllm
CVE-2025-32444Same product: Vllm Vllm
CVE-2025-29783Same product: Vllm Vllm
CVE-2024-11041Same product: Vllm Vllm
CVE-2025-47277Same product: Vllm Vllm
CVE-2025-62164Same product: Vllm Vllm
CVE-2026-34760Same product: Vllm Vllm
CVE-2025-48944Same product: Vllm Vllm
CVE-2026-44222Same product: Vllm Vllm
CVE-2026-44223Same product: Vllm Vllm

Affected Assets

vllm
vllm
≤ 0.7.0

Mitigating Controls

Mitigating Controls (NIST 800-53 r5) AI

Developer testing and evaluation can uncover deserialization flaws before deployment.

Input validation directly stops deserialization of untrusted data by ensuring inputs are valid before processing.

Engineering principles such as safe deserialization and input sanitization structurally prevent the weakness from being introduced.

Integrity verification tools can detect malformed or tampered serialized data after the fact.

Mitigating Controls (NIST CSF 2.0) AI

Derived directly from the weakness types (CWEs) cited in the NVD entry via our AI-authored CWE→CSF cross-walk (authority under review) — links open the control.

PR.PS-02 none match
prevents

PR.PS-02 addresses only post-deployment updates/patching and cannot prevent introduction of unsafe deserialization code, yet it can remediate some instances when the flaw exists in outdated libraries or components.

Mitigating Controls (ISO/IEC 27001:2022 Annex A) AI

Derived directly from the weakness types (CWEs) cited in the NVD entry via our AI-authored CWE→ISO cross-walk (authority under review) — links open the control.

finds

Security testing includes validation of deserialization routines and the use of untrusted data, reducing the likelihood that unsafe object reconstruction will be deployed.

prevents

Requiring vetted libraries, regular updates and SAST before release reduces the likelihood that deserialization logic will accept and act on attacker-controlled serialized objects.

finds

Regular scanning of third-party libraries and timely patching reduce the likelihood that unsafe deserialization vulnerabilities remain active.

none

Mandatory malware scanning of data received over networks or storage media intercepts malicious serialized payloads before they are deserialized by the target application.

References