llama.cpp: Unauthenticated remote code execution in ggml-rpc RDMA transport (CVE-2026-118375)

A stack-based buffer overflow exists in the rdma_recv() function of the ggml-rpc RDMA transport in llama.cpp versions prior to b11865. The vulnerability occurs because the rpc_serve_client() function fails to safely handle data when an attacker advertises RDMA capabilities within a crafted HELLO message. An unauthenticated remote attacker can overflow a 1-byte stack variable with up to 256 KiB of data, leading to a service crash or potential remote code execution.

Impact

Successful exploitation allows for unauthenticated remote code execution in the context of the llama.cpp service, or service denial via application crash.

Recommendation

• Patch llama.cpp by updating to version b11865 or later immediately.
• Disable RDMA capabilities for the ggml-rpc service if network transport is not required in your environment.
• Audit network logs for connections to the llama.cpp RPC port originating from untrusted network segments.