Jackrong-llm-finetuning-guide: MTP weight_map handler and indexed-tensor path traversal in Qwen MTP GGUf pipeline

Two weaknesses were catalogued in the R6410418 Jackrong-llm-finetuning-guide repo (up to commit ef2b17f). CVE-2026-113738 (CVSS 6.3) affects the extract_mtp_heads function of qwen-mtp-gguf/scripts/qwen_mtp_gguf_pipeline.py, in the MTP weight_map handler; CVE-2026-113737 (CVSS 5.5) affects validate_indexed_tensors in the same file, where manipulation of the weight_map argument leads to path traversal (CWE-22). Both are in the Qwen multi-token-prediction (MTP) GGUf conversion/finetuning pipeline.

These are in an LLM finetuning/quantization pipeline (Qwen MTP -> GGUF conversion). If the model weight map or indexed-tensor inputs are attacker-influenced (e.g. a model/weight file supplied from an untrusted hub), the traversal can read or write files outside the intended directory on a finetuning or conversion machine, which is a realistic ML-infrastructure concern for teams converting Qwen MTP models.

Supplying a crafted weight_map / indexed-tensor argument to the conversion script causes path traversal or an uncontrolled manipulation in the MTP weight_map handler.

R6410418/Jackrong-llm-finetuning-guide up to ef2b17f (qwen-mtp-gguf/scripts/qwen_mtp_gguf_pipeline.py)

Validate and constrain weight_map/indexed-tensor paths to the model directory before processing, and only run the conversion pipeline on trusted model inputs.