CVE-2026-114052: vLLM memory exhaustion in MOSS-Audio model serving via /tokenize endpoint

Summary

vLLM versions after 0.24.0 and before 0.31.0 contain a memory exhaustion flaw (CVE-2026-114052) in MOSS-Audio model serving. The component caches a processor for every distinct request-controlled mm_processor_kwargs value in an unbounded dictionary, so unauthenticated attackers can send distinct or oversized mel_config values to the /tokenize endpoint, which bypasses --api-key, to exhaust host memory and crash the API process.

Mitigation

The source does not state a fix yet. Check the original advisory for updates.

Topics

• Inference infrastructure

Related items

• HighCVE-2026-114055: vLLM uncontrolled resource consumption via per-request mm_processor_kwargsSimilar attack · NVD/CVE Database
• HighCVE-2026-114051: vLLM validation bypass in /inference/v1/generate requests with null max_tokensSimilar attack · NVD/CVE Database
• HighCVE-2026-114050: vLLM denial of service via client-supplied multimodal uuid cache desyncSimilar attack · NVD/CVE Database
• HighGHSA-46jr-4wx8-v453: MCP TypeScript SDK: One HTTP request within the default size limit could exhaust server memory during validationSimilar attack · GitHub Advisory Database
• LowLLM Trojans that Increase Response Lengths and CostSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)