unsloth-training

CRITICAL E005: Suspicious download URL detected in skill instructions.

• Suspicious download URL detected (high risk: 0.80). These URLs include direct installer scripts and git-based installs (raw GitHub .py and git+https repo URLs) which are high-risk patterns because piping or installing code from remote/unverified repositories can deliver malware; local endpoints and official package/CDN links are lower risk.

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.30). The runtime workflow in reference/grpo/basic_grpo.py loads and then LLM-ingests GSM8K question text from load_dataset("openai/gsm8k", ...) to form prompt entries used during GRPO training, so outsider-authored free text can be ingested without selecting a specific item.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.70). The skill includes installation commands that fetch and execute remote code (curl https://sh.rustup.rs | sh and wget -qO- https://raw.githubusercontent.com/unslothai/unsloth/main/unsloth/_auto_install.py | python -) and suggests pip installs from git+https://github.com/unslothai/unsloth.git and git+https://github.com/unslothai/unsloth_zoo.git, which are high-confidence remote-code execution vectors if run.