CVE-2026-94626 vllm-project CVE debrief

Who should care

Defenders responsible for systems using vLLM versions up to 0.29.0, especially those with prefill/decode disaggregated deployments, should assess exposure and prioritize patching or mitigating this vulnerability to prevent potential memory exhaustion attacks.

Why it matters

CVE-2026-94626 allows attackers to cause memory exhaustion in vLLM through 0.29.0 by supplying arbitrary tp_size values, impacting systems with prefill/decode disaggregated deployments and requiring prompt patching and monitoring.

• Potential for memory exhaustion and denial of service
• Risk of kernel OOM-kill of the decode worker process
• Need for validation of tp_size parameter in kv_transfer_params
• Potential for attackers to supply arbitrary tp_size values

Technical summary

The vLLM project through version 0.29.0 contains a vulnerability where the tp_size parameter in kv_transfer_params is not validated on OpenAI-compatible completion endpoints. This allows attackers to supply arbitrary tp_size values, potentially leading to unbounded memory allocation and exhaustion in prefill/decode disaggregated deployments, which could trigger kernel OOM-kill of the decode worker process.

Defensive priority

Defenders should prioritize verifying and applying patches for vLLM versions up to 0.29.0, assessing exposure in prefill/decode disaggregated deployments, and monitoring for potential memory exhaustion attacks.

Recommended defensive actions

• Verify and apply patches for vLLM versions up to 0.29.0
• Assess exposure in prefill/decode disaggregated deployments
• Monitor for potential memory exhaustion attacks
• Update to a version of vLLM that validates the tp_size parameter
• Review compensating controls for exposed systems while remediation is scheduled and verified
• Check relevant monitoring, detection, and logs for exposed assets that need extra review
• Track exceptions, retest remediated assets, and close the item only after evidence is documented

Evidence notes

The CVE record and NVD entry provide details on the vulnerability in vLLM through 0.29.0, where the tp_size parameter in kv_transfer_params is not validated, allowing for unbounded memory allocation and potential kernel OOM-kill of the decode worker process.

Sources and references

Methodology and review provenance

AI-assisted synthesis based on stored public vulnerability evidence. System validation, approval state, and publication status do not by themselves establish human review of this revision. PatchSiren helps prioritize defensive review and does not prove exposure or remediation on any system.