CVE-2026-94623 is a denial of service vulnerability in vLLM through version 0.29.0. The issue resides in the NIXL connector's prefix caching mechanism, which does not properly validate block counts when handling multi-prompt completion requests in prefill/decode disaggregated deployments. Attackers can exploit this by sending completion requests containing multiple prompts of different lengths, triggering an assertion failure in the NixlBaseConnectorWorker._apply_prefix_caching function. This causes the decode worker process to terminate unexpectedly and become unavailable until a manual restart is performed.

Successful exploitation results in denial of service by crashing the decode worker component of vLLM, rendering it unavailable until manually restarted. This disrupts the availability of the service relying on the decode worker, potentially impacting dependent applications or users.

Patch status is not yet confirmed — check the vendor advisory for current remediation guidance. Until an official fix is available, avoid sending multi-prompt completion requests with varying prompt lengths in prefill/decode disaggregated deployments or implement monitoring to detect and restart the decode worker if it terminates unexpectedly.