llama.cpp RPC Server Uncaught Exception Denial of Service

Executive Summary

This vulnerability is a denial-of-service (DoS) flaw in the ggml-rpc server of llama.cpp versions b2877 to b11934. It occurs when unauthenticated remote attackers send RPC messages with oversized lengths after the HELLO handshake. The server fails to handle these large sizes properly, triggering an uncaught std::length_error exception that crashes the process, denying service to all clients.

Useful 0 Not Useful 0

Detection Guidance

To detect this vulnerability, monitor for crashes in the ggml-rpc server process after receiving RPC messages. Check server logs for std::length_error exceptions or termination events. Use network traffic analysis tools like Wireshark to inspect RPC message lengths for abnormally large values exceeding std::vector::max_size().

• Check server logs for crashes or exceptions: grep -i 'length_error\|terminate' /var/log/llama-rpc-server.log
• Monitor network traffic for oversized RPC messages: tcpdump -i any -A -n port <rpc_port> | grep -i 'length'

Useful 0 Not Useful 0

Impact Analysis

If you run a vulnerable version of llama.cpp's ggml-rpc server, attackers can remotely crash your server by sending a single malformed message. This disrupts all client connections and services relying on the RPC server, causing downtime and potential data processing interruptions.

Useful 0 Not Useful 0

Mitigation Strategies

Upgrade to llama.cpp version b11935 or later to apply the patch that catches std::length_error exceptions. If immediate upgrade is not possible, restrict network access to the ggml-rpc server port or disable the RPC server until patched.

• Update to b11935: git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp && git checkout b11935
• Temporarily block RPC port (default 8080) in firewall: ufw deny 8080

Useful 0 Not Useful 0