CVE-2024-9052: vLLM deserialization vulnerability in vllm.distributed.GroupCoordinator.recv_object
vLLM's distributed API function `GroupCoordinator.recv_object()` in `vllm/distributed/parallel_state.py` runs `pickle.loads()` on received bytes. If that internal API is exposed to a network, an attacker can run code on the host. The vLLM maintainers dispute the severity: vLLM does not use the code in the way the report describes. The advisory lists no patched version.
Published Mar 20, 2025 · Updated Apr 9, 2025 · Analysis Nov 1, 2026
What an attacker can do
recv_object() receives object bytes from another process in a distributed group and deserializes them with pickle.loads() without checking them. A crafted pickle executes code during loading. The CVSS 3.0 score of 9.8 assumes that an unauthenticated network attacker can send bytes to that function.
The maintainers say the problem exists only if you use these internal APIs in a way that exposes them to a network, as the huntr report assumes. They were not involved in the analysis or the decision to assign the CVE. So the real risk depends on your deployment. A stock vLLM API server does not hand this function to the internet. The risk is real if you run multi-node or distributed setups where the group communication ports are reachable by untrusted hosts, or if you build your own service on GroupCoordinator. In those cases an attacker gets code execution on a GPU host that holds models, keys and internal network access.
Who is affected
The vllm package on PyPI at 0.8.1 and earlier (<= 0.8.1) per the advisory. The advisory text names 0.6.0 as the version in the original report. Check your version:
pip show vllmThen check whether you run distributed inference across hosts, and whether the ports used for group communication are reachable by anything outside your cluster.
vllm
- Ecosystem
- PyPI
- Affected
- <= 0.8.1
- Fixed in
- No patched version yet
- Check
pip show vllm
How to fix it
The advisory lists no patched version. Check vLLM releases newer than 0.8.1 for changes to parallel_state.py. The maintainers' position is that the issue arises only when internal APIs are exposed to a network. The practical mitigation is to keep distributed communication on a private network, firewall the ports between nodes, and never expose these internal APIs to untrusted clients. Do not build network services directly on GroupCoordinator.recv_object().
How to tell if you were hit
Check which hosts connect to your vLLM distributed communication ports and look for any outside your cluster. On vLLM hosts, look for unexpected child processes (shells, curl, wget), new outbound connections and changed files. If you suspect a hit, rotate keys and tokens that the host could read.
Original advisory
From the GitHub Advisory Database (CC BY 4.0).
vllm-project vllm version 0.6.0 contains a vulnerability in the distributed training API. The function vllm.distributed.GroupCoordinator.recv_object() deserializes received object bytes using pickle.loads() without sanitization, leading to a remote code execution vulnerability.
Maintainer perspective
Note that vLLM does NOT use the code as described in the report on huntr. The problem only exists if you use these internal APIs in a way that exposes them to a network as described. The vllm team was not involved in the analysis of this report and the decision to assign it a CVE.
Credited to russellb.
Questions
- Do the vLLM maintainers consider CVE-2024-9052 a vulnerability?
They say vLLM does not use the code as the report describes. The problem exists only if you expose these internal APIs to a network. They were not involved in assigning the CVE.
- Is there a patched vLLM version?
None is listed. The advisory covers 0.8.1 and earlier.
- Is it exploited in the wild?
No source says so and it is not in the KEV catalog. EPSS is low at 0.0033.
References
Read for this analysis:
- https://nvd.nist.gov/vuln/detail/CVE-2024-9052
- https://huntr.com/bounties/ea75728f-4efe-4a3d-9f53-33f2c908e9f8
- https://github.com/vllm-project/vllm/blob/32e7db25365415841ebc7c4215851743fbb1bad1/vllm/distributed/parallel_state.py#L480
- https://github.com/vllm-project/vllm/blob/v0.8.1/vllm/distributed/parallel_state.py#L457
- https://github.com/github/advisory-database/pull/5444
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
