GHSA-ggpf-24jw-3fcw: CVE-2025-24357 Malicious model remote code execution fix bypass with PyTorch < 2.6.0
vLLM's earlier fix for malicious-model code execution (CVE-2025-24357) added `weights_only=True` to `torch.load()`, but that does not protect when PyTorch is older than 2.6.0. vLLM before 0.8.0 installs PyTorch 2.5.1 by default, so loading a malicious model can still run code. Fixed in vLLM 0.8.0.
Published Apr 23, 2025 · Analysis Nov 1, 2026
What an attacker can do
The earlier advisory reported that loading a malicious model could run code on the vLLM host. vLLM's fix passed weights_only=True to torch.load(). The reporters found that this flag was not safe in PyTorch before 2.6.0. PyTorch issued its own advisory for the problem. So the fix did not hold.
The advisory explains why this matters by default. The vLLM requirements file pins PyTorch, and a standard install of vLLM before 0.8.0 pulls in PyTorch 2.5.1. Most users of those vLLM versions therefore ran the unsafe combination. The CWE assigned is dependency on a vulnerable third-party component.
The attacker needs the server to load a model file they crafted. That means a poisoned checkpoint from a public hub, a shared model directory, or a model that a user is allowed to point the server at. The CVSS vector (9.8) assumes network reach and no privileges. Result: code execution as the vLLM process, with the GPUs, other models and any credentials in its environment.
Who is affected
Affected: vLLM before 0.8.0. The issue also depends on PyTorch, since versions of vLLM using PyTorch before 2.6.0 are vulnerable. Check both:
pip show vllm torchA default install of an affected vLLM gets PyTorch 2.5.1.
vllm
- Ecosystem
- PyPI
- Affected
- < 0.8.0
- Fixed in
- 0.8.0
- Check
pip show vllm
How to fix it
Upgrade vLLM to 0.8.0:
pip install --upgrade vllm==0.8.0The advisory's own fix line says to update PyTorch to 2.6.0, so confirm pip show torch reports 2.6.0 or later afterwards. As a habit, load models only from sources you trust, and do not let untrusted users choose which model path the server loads.
How to tell if you were hit
Check which model files your servers loaded and from where. Look for checkpoints that came from unknown publishers, and for pickle-based weight files (.bin, .pt) from untrusted repos. On the host, look for unexpected child processes of the vLLM server, new files and outbound connections right after a model load. The advisory lists no log line.
Original advisory
From the GitHub Advisory Database (CC BY 4.0).
Description
https://github.com/vllm-project/vllm/security/advisories/GHSA-rh4j-5rhw-hr54 reported a vulnerability where loading a malicious model could result in code execution on the vllm host. The fix applied to specify weights_only=True to calls to torch.load() did not solve the problem prior to PyTorch 2.6.0.
PyTorch has issued a new CVE about this problem: https://github.com/advisories/GHSA-53q9-r3pm-6pq6
This means that versions of vLLM using PyTorch before 2.6.0 are vulnerable to this problem.
Background Knowledge
When users install VLLM according to the official manual image
But the version of PyTorch is specified in the requirements. txt file image
So by default when the user install VLLM, it will install the PyTorch with version 2.5.1 image
In CVE-2025-24357, weights_only=True was used for patching, but we know this is not secure. Because we found that using Weights_only=True in pyTorch before 2.5.1 was unsafe
Here, we use this interface to prove that it is not safe. image
Fix
update PyTorch version to 2.6.0
Credit
This vulnerability was found By Ji'an Zhou and Li'shuo Song
Credited to azraelxuemo, russellb.
Questions
- How is this different from CVE-2025-24357?
That advisory reported code execution from loading a malicious model, and vLLM's fix added
weights_only=Truetotorch.load(). This advisory says that fix does not work with PyTorch before 2.6.0.- Does upgrading PyTorch alone fix it?
The advisory says to update PyTorch to 2.6.0, and lists vLLM 0.8.0 as the patched vLLM version. Doing both is the safe route.
- Does it have a CVE number?
This record carries only the GitHub advisory id GHSA-ggpf-24jw-3fcw. PyTorch issued a separate advisory for the underlying
weights_onlyproblem.
References
How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. The analysis is written by an AI model from the sources listed above and checked by automated rules before it is published. Check versions against the vendor's advisory before you act. Methodology
More vLLM vulnerabilities
All vLLM entriesCVE-2026-115911
vLLM: DoS via unbounded chat_template Jinja2 rendering exhausts the renderer thread pool
medium6.5Oct 28, 2026CVE-2026-115912
vLLM: Public multimodal SHM handle forgery reaches worker-side `pickle.loads()` with insecure serialization disabled
high7.5Oct 28, 2026CVE-2026-115913
vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine
medium6.5Oct 28, 2026CVE-2026-115914
vLLM DeepSeek-OCR image processing vulnerable to denial of service
medium6.5Oct 28, 2026CVE-2026-115915
vLLM: Untrusted EPS inputs can trigger denial of service or Ghostscript RCE
high8.1Oct 28, 2026CVE-2026-115916
vLLM: Out-of-vocabulary prompt token id passes input validation and triggers a CUDA device-side assert (engine crash / DoS)
medium6.5Oct 28, 2026
