An unauthenticated remote attacker can cause a denial of service in vLLM versions through 0.29.0 by sending malformed metadata, leading to a complete engine crash.

The vulnerability exists in the NIXL connector, where improper handling of incomplete kv_transfer_params dictionary entries triggers an uncaught KeyError in the EngineCore scheduling logic. This flaw allows an unauthenticated attacker to remotely terminate the decode engine by sending specifically crafted, incomplete metadata requests.

The exploitation of this vulnerability results in a denial of service, rendering the vLLM deployment unavailable for all users. Because the decode engine terminates and requires manual intervention to restart, this represents a significant risk to the availability of AI-driven services, potentially disrupting critical business operations and automated workflows. With a CVSS score of 7.5, the vulnerability is classified as high severity, reflecting the ease of exploitation and the impact on system uptime.

Immediate Action: Monitor official repository channels for the release of version 0.29.1 or higher and apply the update immediately upon availability.

Proactive Monitoring: Review system logs for recurring "KeyError" exceptions within the EngineCore scheduling component and monitor for unexpected service terminations of the vLLM decode engine.

Compensating Controls: Implement network-level access controls or a Web Application Firewall (WAF) to restrict access to the NIXL connector endpoints to known, trusted internal IP addresses, thereby reducing the attack surface.

Public Exploit Available: No (exploit_available: false)

Given the high severity and the potential for service disruption, administrators should prioritize the application of vendor patches as soon as they are released. In the interim, isolating the NIXL connector from untrusted network segments is essential to prevent unauthorized access and potential exploitation of this denial of service vulnerability.