An unauthenticated remote attacker can trigger a denial of service in vLLM by supplying arbitrary tp_size values, leading to memory exhaustion and process termination.

This is a memory allocation vulnerability (CWE-789) occurring in OpenAI-compatible completion endpoints. An unauthenticated attacker can supply malicious input to the tp_size parameter within kv_transfer_params, forcing the application to allocate unbounded memory.

Successful exploitation of this vulnerability results in a denial of service by triggering a kernel OOM-kill of the decode worker process. This leads to system downtime for the affected LLM inference service, disrupting dependent applications and potentially causing significant operational impact for organizations relying on vLLM for real time request processing. The high CVSS score of 7.5 reflects the critical impact on service availability.

Immediate Action: Monitor for the release of a patched version of vLLM and update immediately once available; no specific fix version is currently identified.

Proactive Monitoring: Review access logs for anomalous requests to completion endpoints, specifically monitoring for unusual values within the kv_transfer_params payload.

Compensating Controls: Implement a Web Application Firewall (WAF) or API gateway to validate incoming requests and restrict access to completion endpoints to authorized traffic only, effectively blocking untrusted input.

Public Exploit Available: Unknown

Given the potential for service disruption via remote denial of service, this vulnerability presents a significant risk to infrastructure stability. Administrators should prioritize restricting network access to the affected endpoints and remain vigilant for security patches from the vLLM project, applying them as soon as they are made available to mitigate the risk of memory exhaustion attacks.