hugging-face-model-trainer

SUSPICIOUS: The skill is largely coherent with its stated Hugging Face training purpose and uses official HF/Astral tooling, so it is not malicious. However, it enables autonomous paid job submission and execution of remote training scripts with HF_TOKEN available in the runtime, creating meaningful supply-chain and cost/credential exposure if unpinned third-party scripts are used.

No direct evidence of intentional malware or covert data theft is present in this module. The security risk is dominated by supply-chain and code-execution surfaces: it executes code from a runtime-cloned llama.cpp repository (including pip installs and native builds without pinning) and it may execute remote Hugging Face model/tokenizer repository code when TRUST_REMOTE_CODE is enabled. Treat this as a high supply-chain risk tool unless you fully control and pin inputs (model repos, llama.cpp revision) and keep TRUST_REMOTE_CODE disabled.