vLLM Denial of Service via P2P KV Offloading Misconfiguration

Executive Summary

vLLM versions up to 0.29.0 have a denial of service vulnerability in P2P KV offloading. When OffloadingConnector is set with TieringOffloadingSpec and a peer-to-peer secondary tier, attackers can provide fake remote host and port values in kv_transfer_params. This creates unreachable peer sessions that keep ZeroMQ sockets open until the context quota is exhausted, leading to a ZMQError that crashes EngineCore and halts all inference operations.

Useful 0 Not Useful 0

Detection Guidance

This vulnerability can be detected by monitoring for unreachable peer sessions in vLLM's P2P KV offloading. Check for ZeroMQ socket exhaustion in EngineCore logs and verify if OffloadingConnector is configured with TieringOffloadingSpec. Look for ZMQError crashes during inference operations.

Useful 0 Not Useful 0

Impact Analysis

This vulnerability can cause your vLLM-based inference system to crash, stopping all ongoing and new inference tasks. It results in downtime and service disruption, potentially affecting applications relying on real-time AI model outputs.

Useful 0 Not Useful 0

Compliance Impact

This vulnerability causes service disruption by crashing the inference engine, which could lead to unauthorized data access or processing delays. Such disruptions may violate availability requirements in GDPR and HIPAA, potentially resulting in compliance failures.

Useful 0 Not Useful 0

Mitigation Strategies

Upgrade vLLM to a version beyond 0.29.0 to address the denial of service vulnerability in P2P KV offloading. Ensure OffloadingConnector configurations do not use TieringOffloadingSpec with arbitrary remote host and port values in kv_transfer_params.

Useful 0 Not Useful 0