Executive Summary
A critical vulnerability (CVE-2026-118375) has been identified in llama.cpp version b8841 before b11865. This out-of-bounds write vulnerability in the ggml-rpc RDMA transport rdma_recv() function allows unauthenticated peers to overflow the rpc-server stack, potentially leading to code execution or server crashes. The vulnerability has a CVSS score of 8.1, indicating high severity.
Technical Analysis
The vulnerability is classified as an out-of-bounds write issue in the ggml-rpc RDMA transport rdma_recv() function. The attack vector involves unauthenticated peers advertising RDMA capabilities in the HELLO message and sending up to 256 KiB into a 1-byte stack variable in rpc_serve_client(). This can cause a stack overflow, potentially leading to code execution or server crashes. The root cause is a lack of proper input validation in the rdma_recv() function.
How It Gets Exploited
An unauthenticated remote attacker on the same network can exploit this vulnerability. The attacker would advertise RDMA capabilities in the HELLO message and send a crafted payload to the rpc-server, specifically targeting the rpc_serve_client() function. The payload would overflow the 1-byte stack variable, causing a stack-based buffer overflow. This would allow the attacker to potentially execute code or crash the server.
Impact Assessment
The affected product is llama.cpp version b8841 before b11865. An attacker can achieve arbitrary code execution or cause a denial-of-service (DoS) condition. The CVSS score of 8.1 indicates a high severity level, with high impacts on confidentiality, integrity, and availability.
Recommended Actions
To mitigate this vulnerability, update llama.cpp to version b11865 or later. Additionally, implement network segmentation and firewalls to restrict access to the affected service. Monitor network traffic for suspicious activity and implement intrusion detection systems to detect potential exploitation attempts.
Sources
- National Vulnerability Database (NVD)
- Vulncheck