CVE-2026-94626

Summary

The vLLM library fails to validate the tp_size parameter in kv_transfer_params, allowing unbounded memory allocation during prefill/decode operations. This vulnerability can lead to memory exhaustion and kernel OOM-kill of decode worker processes.

Exploitability

Exploitation is straightforward as attackers can simply supply arbitrary tp_size values through OpenAI-compatible completion endpoints without requiring authentication or user interaction.

Blast radius

The vulnerability affects prefill/decode disaggregated deployments where attackers can trigger memory exhaustion across multiple decode worker processes, potentially causing service disruption and denial of service.

Detection

Monitor for kernel OOM-kill messages in system logs when decode worker processes are terminated, and watch for unusually high memory consumption patterns during prefill/decode operations.

Prioritized remediation

Upgrade to vLLM version 0.29.1 or later to address the tp_size parameter validation issue.

Analysis generated locally by garnet-coder (no data left the box). AI-assisted — verify against primary sources before acting.