CVE-2026-118375: llama.cpp b8841 before b11865 Stack-based Buffer Overflow via ggml-rpc RDMA rdma_recv()
llama.cpp b8841 before b11865 contains an out-of-bounds write vulnerability in the ggml-rpc RDMA transport rdmarecv() function that allows unauthenticated peers to overflow the rpc-server stack. Attackers advertising RDMA capabilities in the HELLO message can send up to 256 KiB into a 1-byte stack variable in rpcserveclient(), crashing the server or potentially executing code.
Affected Software
Event History
Frequently Asked Questions
What access does an attacker need to exploit this issue?
An attacker only needs network reachability to the rpc-server and does not need authentication or user interaction. Exploitation requires the attacker to advertise RDMA capabilities in the HELLO message and then send crafted RDMA data.
Which deployments are exposed?
Deployments running llama.cpp versions from b8841 up to, but not including, b11865 are affected when the ggml-rpc server is reachable by untrusted peers. The vulnerable path is the ggml-rpc RDMA transport.
What can happen if exploitation succeeds?
A malicious peer can write up to 256 KiB into a 1-byte stack variable in rpc_serve_client(). This can crash the rpc-server and may enable code execution.