AI/ML: llama.cpp WebUI injection flaw and Transformers remote patch claim

Security discussion dominated the day's AI and ML open source activity, centered on a reported WebUI flaw and a low-credibility remote patch offer. Both items involve popular inference and model libraries that developers rely on locally.

llama.cpp WebUI query-parameter injection

A security report on the ggml-org/llama.cpp repository described an instruction injection path through the Query parameter of the llama-server WebUI. The issue is framed as a CSRF-style flaw that can enable prompt injection and, when tools are enabled, remote code execution. Users running the WebUI exposed to untrusted input have a direct reason to review access controls and pending fixes.

Remote script offered as Transformers v5.x patch

A single post on the huggingface/transformers repository presented a one-line PowerShell command that fetches and runs a remote script, claiming to resolve head_dim crashes, out-of-memory errors, and tokenizer bugs. The approach of pulling unsigned code from outside the project raises immediate supply-chain concerns for anyone maintaining local Transformers installs.