CVE-2025-62164: vLLM memory corruption via unvalidated prompt-embedding tensors

# CVE-2025-62164: vLLM memory corruption via unvalidated prompt-embedding tensors

https://gcdradar.vercel.app/incidents/cve-2025-62164-vllm-memory-corruption-via-unvalidated-prompt · AIR-2025-0521 · closed 2025-11-20

- Lane: Patch now
- Severity: 2 (Disclosed vulnerability)
- Fix status: Fix status unknown
- First reported: 2025-11-20
- AI role: Attack on AI
- Products: vLLM
- CVE ids: CVE-2025-62164
- CVSS: 8.8

## Summary

vLLM from 0.10.2 before 0.11.1 loaded user-supplied prompt embeddings in its Completions API with torch.load() and no sparse tensor validation, so a crafted tensor could write out of bounds during to_dense(). The flaw crashes the server and may allow remote code execution; 0.11.1 fixes it.

## What to do

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

## Latest updates

- 2025-11-21: CVE-2025-62164: vLLM is an inference and serving engine for large language models (LLMs). (https://nvd.nist.gov/vuln/detail/CVE-2025-62164)
- 2025-11-20: CVE-2025-62164: vLLM deserialization vulnerability leading to DoS and potential RCE (https://github.com/advisories/GHSA-mrw7-hf4f-83pf)

## Sources

- 2025-11-20 GitHub Advisory Database: CVE-2025-62164: vLLM deserialization vulnerability leading to DoS and potential RCE https://github.com/advisories/GHSA-mrw7-hf4f-83pf
- 2025-11-21 NVD: CVE-2025-62164: vLLM is an inference and serving engine for large language models (LLMs). https://nvd.nist.gov/vuln/detail/CVE-2025-62164

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