vLLM versions up to 0.29.0 are susceptible to a denial‑of‑service flaw arising from how the OffloadingConnector handles peer‑to‑peer KV offloading. When configured with a TieringOffloadingSpec and a second‑tier that uses p2p, an attacker can supply arbitrary remote host and port values in the kv_transfer_params. This forces the system to attempt to establish a ZeroMQ connection to an unreachable host, leaving the socket open until the internal context quota runs out. The resulting uncaught ZMQError crashes EngineCore, halting all inference operations.

The vulnerability affects the vllm‑project vllm library for all operating environments0 release. Users running vLLM built with the OffloadingConnector and a peer‑to‑peer secondary tier are at risk. Any deployment that imports vllm v0.29.0 (or earlier) may be impacted, regardless of platform, because the vulnerability resides within the library itself.

Because the flaw is triggered by externally supplied network parameters, an attacker with network reach to a target application can exploit it without authentication. The closely‑packed socket allocations can exhaust the ZeroMQ context quota quickly, leading to a crash. The CVSS score of 8.7 reflects high severity, and the EPSS score of 0.00446 indicates a very low but non‑zero exploitation probability, so while the risk is moderate, the flaw remains significant. As the bug exists in vLLM and was publicly disclosed in 2026, attackers could already try to feed malformed kv_transfer_params to cause repeated crashes, potentially resulting in denial of service for the inference service.