CVE-2025-71379 vllm CVE debrief

Who should care

Organizations using vLLM versions >= 0.6.3 and < 0.9.0 should prioritize patching to prevent potential denial-of-service attacks. This vulnerability could impact services relying on vLLM for critical operations, especially those with high availability requirements. Developers and administrators of applications built on top of vLLM should assess their exposure and take appropriate measures.

Technical summary

The vulnerability exists in multiple components of vLLM, including vllm/lora/utils.py, the phi4mini tool parser, and the OpenAI-compatible serving chat endpoint. Specifically, the regex patterns in these components are susceptible to catastrophic backtracking when processing crafted input with nested or repeated structures. This can lead to severe CPU consumption and performance degradation, effectively causing a denial of service. The CVSS score for this vulnerability is 5.3, indicating a medium severity level.

Defensive priority

Medium priority due to potential service disruption with CVSS score of 5.3.

Recommended defensive actions

• Inventory vLLM installations to identify potentially affected versions.
• Review official advisories and documentation for patching guidance.
• Apply vendor-supported remediation by updating to a patched version (>= 0.9.0).
• Implement compensating controls such as input validation and rate limiting.
• Monitor for suspicious activity and performance degradation.

Evidence notes

The primary evidence for this vulnerability comes from the CVE record and NVD detail pages. The vulnerability affects vLLM versions >= 0.6.3 and < 0.9.0. Defenders should verify the version of vLLM in use and check for official patches or advisories from the vendor. The CVE record and NVD detail provide additional information on the vulnerability's characteristics and potential impact.

Official resources