The vulnerability resides in the NIXL connector's metadata handler for prefill/decode disaggregated deployments in vLLM. An attacker can supply requests that omit required keys in the kv_transfer_params dictionary, causing the EngineCore scheduling logic to raise an uncaught KeyError. The error propagates through the decode engine, terminating the worker process and rendering all routed requests unusable until the affected service is manually restarted. The weakness is classified as CWE-248.

The issue affects all instances of the vllm-project vllm application with versions through 0.29.0. No specific sub‑component versions are delineated beyond the stated upper bound of 0.29.0.

The CVSS score of 8.7 indicates a high severity of impact. EPSS data is unavailable, so the likelihood of exploitation cannot be quantified but the lack of a KEV listing suggests no known public exploits. The likely attack vector is external, inferred from the description that an attacker can send crafted requests to trigger the failure. An adversary with network access to the vLLM endpoint could therefore disrupt service continuity until a manual restart occurs.