A critical denial of service vulnerability in vLLM versions 0.29.0 and earlier allows unauthenticated attackers to crash the inference engine by exploiting improper resource management in P2P KV offloading.

This vulnerability, identified as CWE-770 (Allocation of Resources Without Limits or Throttling), occurs when the OffloadingConnector is configured with a peer-to-peer secondary tier. An unauthenticated attacker can supply arbitrary remote host and port parameters within kv_transfer_params to initiate unreachable peer sessions, which exhaust system resources by leaking ZeroMQ sockets until the engine crashes.

Successful exploitation of this vulnerability results in a complete denial of service for the affected vLLM instance. Because the crash impacts the EngineCore, all active inference tasks will be terminated, leading to significant service disruption and potential loss of data in flight. Given the CVSS score of 7.5, this high-severity flaw poses a direct threat to the availability of AI-driven services relying on the vLLM framework.

Immediate Action: Update vLLM to a version beyond 0.29.0 as soon as a patch is released by the vendor.

Proactive Monitoring: Monitor system logs for repeated or anomalous ZMQError entries and unexpected connection attempts to internal P2P offloading endpoints.

Compensating Controls: Implement strict network ingress filtering to ensure that only trusted peer nodes can communicate with the vLLM P2P offloading interfaces, effectively restricting access to the vulnerable parameters.

Public Exploit Available: No (exploit_available: false)

This vulnerability represents a significant availability risk for organizations deploying vLLM. Administrators must prioritize upgrading their software to the latest secure version immediately upon availability. In the interim, ensure that the P2P KV offloading feature is strictly isolated within your network architecture to prevent unauthorized access to the vulnerable control interfaces.