CVE-2026-94626: vLLM input validation bypass in kv_transfer_params

Executive brief

vLLM, an open-source inference engine for large language models, fails to validate the tp_size parameter in OpenAI-compatible API endpoints. An attacker can send malicious requests with arbitrary tp_size values to exhaust server memory, causing the decode worker process to crash and creating a denial-of-service condition that disrupts LLM service availability.

Technical details

The tp_size parameter in kv_transfer_params endpoint lacks input validation, allowing an attacker to supply unbounded values that trigger excessive memory allocation. This affects disaggregated prefill/decode deployments where attackers with network access to the OpenAI-compatible completion endpoint can craft requests to exhaust memory and trigger kernel OOM-kill of worker processes. No authentication is required to exploit this vulnerability.

Affected products

• vLLM Project vLLM through 0.29.0

Timeline

• 2026-09-21: disclosed

References

• https://github.com/vllm-project/vllm
• https://github.com/vllm-project/vllm/blob/v0.29.0/vllm/distributed/kv_transfer/kv_connector/utils.py