Executive brief
llama.cpp is a C++ framework for running large language model inference. When started with the --reranking flag to enable document reranking functionality, the API fails to validate input parameters and crashes on malformed requests. A remote attacker can send a specially crafted POST request with a negative "top_n" value to cause the service to return HTTP 500 errors and exhaust memory, making the inference service unavailable to legitimate users.
Technical details
The vulnerability is an integer handling error (CWE-190) in the /rerank endpoint handler, specifically in tools/server/server-context.cpp (line 4048) where the top_n parameter is extracted from the JSON request body without validation. When a negative top_n value is passed, it triggers a std::bad_alloc exception in the memory allocation routines of format_response_rerank (tools/server/server-common.cpp, line 1245), causing the server to return HTTP 500 errors. The attack requires the server to be explicitly started with the --reranking flag (not enabled by default) and network access to the HTTP API endpoint. The vulnerability is 100% reproducible with negative top_n values and causes denial of service through repeated crashes, though the server process survives individual incidents. No patch has been released as of the latest stable version.
Affected products
- ggml llama.cpp through commit 97f06e9
Timeline
- 2026-08-12: disclosed
- 2026-09-01: advisory: Published to NVD