CVE-2025-29783: vLLM Allows Remote Code Execution via Mooncake Integration
vLLM with the Mooncake KV-transfer integration unpickles raw network bytes received over ZMQ/TCP. An attacker who can reach the socket can run code on the distributed vLLM hosts. Versions 0.6.5 up to 0.8.0 are affected. Upgrade to vLLM 0.8.0.
Published Mar 19, 2025 · Updated Jul 2, 2025 · Analysis Nov 1, 2026
What an attacker can do
The Mooncake pipe in vllm/distributed/kv_transfer/kv_pipe/mooncake_pipe.py moves KV cache between hosts for disaggregated prefilling. recv_tensor() calls _recv_impl(), which passes the received bytes straight to pickle.loads(). Pickle can run arbitrary code while loading, so a crafted payload executes on the receiving host.
The advisory notes there are no network or authentication controls to stop an arbitrary sender from reaching the service. The socket is exposed over the network by design. The receiver connects to the decode host address set in mooncake.json, and an attacker positioned to send to those sockets can deliver the payload. The CVSS vector uses adjacent network access with low privileges.
A hit gives code execution with the rights of the vLLM process on a GPU host, which usually holds model weights, API keys and access to other cluster nodes. The scope is changed, so one compromised node can lead to others. Only deployments that enable the Mooncake integration are exposed. Standard single-node vLLM serving does not use this pipe.
Who is affected
vllm (pip) from 0.6.5 up to, but not including, 0.8.0, and only where Mooncake KV transfer is configured (a mooncake.json for distributed prefill and decode). Check your version:
pip show vllmAlso check whether your launch config sets a Mooncake KV connector.
vllm
- Ecosystem
- PyPI
- Affected
- >= 0.6.5, < 0.8.0
- Fixed in
- 0.8.0
- Check
pip show vllm
How to fix it
Upgrade to vllm 0.8.0:
pip install --upgrade "vllm>=0.8.0"The advisory says the issue is resolved by pull request 14228 (commit 288ca110). If you cannot upgrade, do not use the Mooncake integration, and restrict the ZMQ/TCP ports to the trusted hosts in your prefill and decode setup with firewall rules.
How to tell if you were hit
Check the prefill and decode hosts for connections to the Mooncake ZMQ ports from addresses that are not in your mooncake.json. Look for unexpected child processes of the vLLM worker, new files, or outbound connections from it. Review the firewall logs for traffic to those ports from outside the cluster.
Original advisory
From the GitHub Advisory Database (CC BY 4.0).
Summary
When vLLM is configured to use Mooncake, unsafe deserialization exposed directly over ZMQ/TCP will allow attackers to execute remote code on distributed hosts.
Details
- Pickle deserialization vulnerabilities are well documented.
- The mooncake pipe is exposed over the network (by design to enable disaggregated prefilling across distributed environments) using ZMQ over TCP, greatly increasing exploitability. ~~Further, the mooncake integration opens these sockets listening on all interfaces on the host, meaning it can not be configured to only use a private, trusted network.~~
Only sender_socket and receiver_ack are allowed to be accessed publicly, while the data actually decompressed by pickle.loads() comes from recv_bytes. Its interface is defined as self.receiver_socket.connect(f\"tcp://{d_host}:{d_rank_offset + 1}\"), where d_host is decode_host, a locally defined address 192.168.0.139,from mooncake.json (https://github.com/kvcache-ai/Mooncake/blob/main/doc/en/vllm-integration-v0.2.md?plain=1#L36).
- The root problem is `recv_tensor()` calls `_recv_impl` which passes the raw network bytes to
pickle.loads(). Additionally, it does not appear that there are any controls (network, authentication, etc) to prevent arbitrary users from sending this payload to the affected service.
Impact
This is a remote code execution vulnerability impacting any deployments using Mooncake to distribute KV across distributed hosts.
Remediation
This issue is resolved by https://github.com/vllm-project/vllm/pull/14228
Credited to JosephTLucas, russellb, kexinoh.
Questions
- Is every vLLM server affected by CVE-2025-29783?
No. Only deployments that use the Mooncake integration for distributed KV transfer, on versions 0.6.5 up to 0.8.0, are affected.
- What fixes it?
vLLM 0.8.0, which includes the change from pull request 14228.
- Does the attacker need to be on the same network?
The CVSS vector is adjacent network with low privileges. The attacker must be able to send data to the Mooncake ZMQ sockets, so network placement and firewalling decide exposure.
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
