freenode

← digests

Security fixes land in llama.cpp and PyTorch

AI & ML2026-11-16

Two security patches surfaced in major ML runtimes. One targets an integer overflow in llama.cpp's RPC path, and the other tightens path checks in a PyTorch storage reader.

llama.cpp patch targets ggml_nbytes overflow

A pull request on ggml-org/llama.cpp proposes a fix for an integer overflow in ggml_nbytes() in the ggml-rpc code, described as leading to an out-of-bounds read or write. The reporter has asked that the proposed fix be assigned a CVE. Developers who ship llama.cpp RPC should track the change, though the thread does not yet record a merged fix or an assigned identifier.

PyTorch weight map path check proposed

A pull request on pytorch/pytorch aims to validate weight_map file paths in QuantizedHuggingFaceStorageReader, framed as a security fix under a placeholder CVE identifier. The submission is noted as missing a required actionable issue link. Readers should treat the CVE number as provisional until the project records a proper issue and a confirmed identifier.