model-trainer

Warn

Audited by Socket on Mar 28, 2026

3 alerts found:

Anomalyx3
AnomalyLOW
skills/model-trainer/scripts/convert_to_gguf.py

This script is a legitimate automation for converting and uploading LoRA-merged Hugging Face models to GGUF and quantized formats. It does not itself contain explicit malicious payloads, hidden backdoors, or obfuscated code. However it performs multiple high-risk supply-chain and execution actions: it loads remote model/tokenizer code with trust_remote_code=True, clones and executes scripts from an external GitHub repo, installs packages and builds binaries, and uploads model artifacts to Hugging Face. These behaviors create significant supply-chain and exfiltration risk if any of the external repositories or credentials are compromised or if it is run in an environment with sensitive data or shared /tmp. I assess low probability that this script is intentionally malicious, but the security risk is moderate-to-high due to execution of untrusted code and network artifact uploads — it should only be run in a controlled, isolated environment after auditing the external repositories and ensuring tokens/credentials are safe.

Confidence: 90%Severity: 60%
AnomalyLOW
SKILL.md

SUSPICIOUS: The skill is largely coherent with its stated Hugging Face training purpose and mostly uses official same-ecosystem services, so it does not look malicious. However, it enables autonomous remote job submission, forwards HF_TOKEN to remote runtimes, and permits arbitrary remote script URLs, making it a medium-risk skill with notable supply-chain and real-world action concerns.

Confidence: 85%Severity: 56%
AnomalyLOW
skills/model-trainer/SKILL.md

SUSPICIOUS: The skill is largely coherent with its stated Hugging Face training purpose and mostly uses official same-ecosystem services, so it does not look malicious. However, it enables autonomous remote job submission, forwards HF_TOKEN to remote runtimes, and permits arbitrary remote script URLs, making it a medium-risk skill with notable supply-chain and real-world action concerns.

Confidence: 85%Severity: 56%
Audit Metadata
Analyzed At
Mar 28, 2026, 08:19 PM
Package URL
pkg:socket/skills-sh/aisa-group%2Fskill-inject%2Fmodel-trainer%2F@a8bd5833157b6873a841122bd72a44971e1fb5ab