llama.cpp contributor seeks CVE for ggml size overflow

A contributor to llama.cpp has asked maintainers whether an integer overflow in the library's tensor size calculation qualifies for a CVE and an official security advisory.

omarrrfds wrote to Georgi Gerganov that the flaw is in ggml_nbytes(), which computes how many bytes a tensor occupies. An overflow there can make later code treat a buffer as smaller than the data it handles. In the RPC path, the author says, that mismatch can read or write outside allocated memory. llama.cpp is widely used to run large language models on local machines, so a memory-safety bug in that size calculation would matter to anyone exposing the RPC interface.

The author said maintainers had already reviewed a proposed fix, and asked whether it is accepted from a security standpoint or whether more work is required before it can be merged. The request also seeks researcher credit. It does not include a CVE identifier, a maintainer decision, or confirmation that the change has landed.