llama.cpp WebUI injection and PyTorch decode kernels

Security and kernel work led the day in open AI infrastructure. A llama.cpp WebUI injection report sat alongside PyTorch Inductor changes aimed at decode shapes, with smaller GGUF and signing notes trailing.

Instruction injection via llama.cpp WebUI query parameter

A security issue was filed against ggml-org/llama.cpp describing instruction injection through the WebUI Query parameter (?q=). The report frames it as a CSRF or prompt-injection flaw in llama-server that can trigger tool execution. Operators exposing the WebUI need to treat untrusted query strings as a path to unintended model actions.

Inductor auto-routes small-M decode shapes to swap_ab NVGEMM

A pytorch/pytorch pull request proposes that Inductor automatically send small-M decode shapes (M<=64) to swap_ab NVGEMM kernels. The goal is parity with routing already used in vLLM and SGLang. Stacks that compile decode GEMMs through Inductor may obtain more consistent short-sequence performance on NVIDIA hardware.

GGUF writer methods extended for prefix-LM

A small ggml-org/llama.cpp thread adds methods to gguf_writer. One participant extended the change for prefix-LM workflows and reported matching evaluation numbers. The work remains incremental serialization tooling rather than a model-format shift.