slime

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

• Third-party content exposure detected (medium risk: 0.30). slime ingests outsider-authored free text at runtime via user-provided training/evaluation JSONL prompts in the required workflow (e.g., --prompt-data /path/to/data.jsonl with prompt and optional chat content), and those prompt strings are then fed into the SGLang rollout generation.

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

• Potentially malicious external URL detected (high risk: 0.90). The SKILL instructs users to run "git clone https://github.com/THUDM/slime.git" and then install/run the fetched code (pip install -r / pip install -e .), which fetches and executes remote repository code as part of setup, so the GitHub URL is a runtime/execution dependency.