CVE-2025-32444: vLLM Remote Code Execution via Mooncake Integration

# CVE-2025-32444: vLLM Remote Code Execution via Mooncake Integration

https://gcdradar.vercel.app/incidents/cve-2025-32444-vllm-remote-code-execution-via-mooncake-integ · AIR-2025-0277 · closed 2025-04-29

- Lane: Patch now
- Severity: 2 (Disclosed vulnerability)
- Fix status: Fix status unknown
- First reported: 2025-04-29
- AI role: Attack on AI
- Products: vLLM
- CVE ids: CVE-2025-32444
- CVSS: 10

## Summary

vLLM instances utilizing the Mooncake integration were vulnerable to remote code execution due to insecure pickle-based serialization over ZeroMQ sockets. The sockets were bound to all interfaces, increasing exposure to network attackers.

## What to do

No action recorded; read the sources below and the vendor's advisory.

## Latest updates

- 2025-04-30: CVE-2025-32444: vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. (https://nvd.nist.gov/vuln/detail/CVE-2025-32444)
- 2025-04-29: CVE-2025-32444: vLLM Vulnerable to Remote Code Execution via Mooncake Integration (https://github.com/advisories/GHSA-hj4w-hm2g-p6w5)

## Sources

- 2025-04-29 GitHub Advisory Database: CVE-2025-32444: vLLM Vulnerable to Remote Code Execution via Mooncake Integration https://github.com/advisories/GHSA-hj4w-hm2g-p6w5
- 2025-04-30 NVD: CVE-2025-32444: vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. https://nvd.nist.gov/vuln/detail/CVE-2025-32444

Verify against the vendor's advisory before changing production. Data CC BY-SA 4.0, https://gcdradar.vercel.app.