Overview
The vulnerability CVE-2025-47277 is a highly critical security flaw found in the vLLM, an inference and serving engine for large language models. Specifically, the vulnerability affects the `PyNcclPipe` KV cache transfer integration with the V0 engine in versions 0.6.5 through 0.8.4. The exploitation of this vulnerability could potentially lead to system compromise or data leakage, posing significant risks to any environment using the affected configurations. The severity of this vulnerability underscores the importance of immediate action to mitigate the risks.
Vulnerability Summary
CVE ID: CVE-2025-47277
Severity: Critical (CVSS: 9.8)
Attack Vector: Network
Privileges Required: None
User Interaction: None
Impact: System Compromise and Data Leakage
Affected Products
Share secrets securely
Ameeba is private infrastructure for communication and sensitive work built on encrypted identity instead of exposed corporate identity systems.
Passwords, credentials, confidential files, screenshots, internal discussions, sensitive AI context, and private coordination should not become exposed across ordinary communication platforms.
- • Encrypted identity
- • Private Spaces for organizations and teams
- • End-to-end encrypted chat, calls, files, and notes
- • Sensitive AI work and protected collaboration
- • Built for information that cannot leak
Our mission is to secure human work alongside AI.
Product | Affected Versions
vLLM | 0.6.5 to 0.8.4
How the Exploit Works
The vulnerability arises from vLLM’s implementation of the `PyNcclPipe` class to establish a P2P communication domain for data transmission between distributed nodes. The `TCPStore` interface, part of the PyTorch framework, was intended to listen only on the IP address specified by the `–kv-ip` CLI parameter, thus limiting exposure to a private, secured network. However, due to a security oversight, the `TCPStore` interface listens on all interfaces, disregarding the provided IP address. This behavior exposes the vLLM engine to potential unauthorized access and exploitation over the network.
Conceptual Example Code
Here’s a conceptual example of an HTTP request that could exploit this vulnerability:
POST /vLLM/execute HTTP/1.1
Host: vulnerable.example.com
Content-Type: application/json
{
"command": "dump_kvcache",
"kvstore": "PyNcclPipe"
}In this conceptual example, an attacker sends a malicious `POST` request to the vLLM engine’s execution endpoint. The `”dump_kvcache”` command could force the vLLM engine to dump the contents of the KV cache, potentially leaking sensitive data over the network.
Mitigation Steps
This vulnerability has been patched in vLLM version 0.8.5, which now limits the `TCPStore` socket to the private interface as configured. All users are advised to update to this version or later to mitigate the risks posed by CVE-2025-47277.
In situations where immediate patching is not feasible, deploying a Web Application Firewall (WAF) or Intrusion Detection System (IDS) could serve as a temporary mitigation measure. These systems can monitor and control incoming network traffic based on predefined security policies, potentially preventing exploitation of this vulnerability.
Despite these measures, it is imperative to apply the vendor patch as soon as possible to fully address the vulnerability.