An out‑of‑bounds memory write is triggered when the llama.cpp server parses the n_discard parameter from JSON input without verifying that it is non‑negative. Supplying a negative value causes the memory removal routine to use a reversed range and negative offset, corrupting memory during the token evaluation loop. This deterministic corruption can crash the process or, in a remote context, enable code execution.

The vulnerability affects the ggml-org llama.cpp implementation, specifically versions associated with commit 55d4206c8 and earlier. No version list is provided, so any release prior to a forthcoming fix is potentially exposed. The affected component is the llama.cpp server’s completion endpoints receiving JSON input.

The CVSS score of 8.8 indicates a high severity, but the EPSS score of less than 1% suggests a low exploitation probability at this time. The vulnerability is not listed in the KEV catalog. A remote attacker can trigger the flaw by sending a crafted JSON request to the server’s completion endpoint with a negative n_discard value. Without a patch, mitigating such an attack would require input validation or network restriction.