The update addresses out-of-bounds (OOB) reads in the UGM tokenizer's precompiled_charsmap handling, which could lead to heap-buffer-overflows from malicious T5/UGM GGUF files. The fix replaces unsafe `strlen()` with a bounds-checked loop and validates minimum blob sizes before reading.
A critical vulnerability in a popular local LLM inference engine is patched, reducing the attack surface for users running untrusted GGUF models.