CVE-2024-42479: 
Homebrew vulnerability analysis and mitigation

CVE-2024-42479 is a critical security vulnerability discovered in llama.cpp, a C/C++ library for LLM inference. The vulnerability stems from an unsafe data pointer member in the rpc_tensor structure that can cause arbitrary address writing. This vulnerability was discovered by 7resp4ss from 360 Vulnerability Research Institute and was fixed in version b3561 (GitHub Advisory, NVD).

The vulnerability is classified as a "Write-what-where" condition in the rpc_server::set_tensor function, with a CVSS v3.1 score of 9.8 (Critical). The issue arises from the unsafe handling of data pointers within the rpc_tensor structure, which can be controlled by users to write data to arbitrary memory locations. The vulnerability chain involves multiple function calls, including start_rpc_server, rpc_serve_client, and ultimately ggml_backend_cpu_buffer_set_tensor where the arbitrary write occurs (Security Online, GitHub Advisory).

The vulnerability has severe implications as it can be exploited to achieve Remote Code Execution (RCE). The issue affects llama_cpp_python, a Python binding for the llama.cpp library, which has over 3 million downloads, significantly amplifying the potential impact. When successfully exploited, this vulnerability could lead to unauthorized access, data breaches, and potentially full system compromise (Security Online).

A proof-of-concept exploit has been developed and demonstrated by security researchers. The vulnerability can be combined with another arbitrary address reading vulnerability to achieve Remote Code Execution. A demonstration video showing the successful exploitation has been made available by the researchers (GitHub Advisory).

Users are strongly advised to update to version b3561 or later of llama.cpp to address this vulnerability. Organizations using llama_cpp_python in their AI and machine learning projects should prioritize this update due to the critical nature of the vulnerability (ASEC).