model-merging
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill describes a workflow for benchmarking merged models that involves processing external prompts and model outputs which could be leveraged for injection attacks.\n
- Ingestion points: Evaluation scripts in references/evaluation.md and SKILL.md ingest user-defined test prompts and model-generated responses.\n
- Boundary markers: None explicitly defined in the provided code snippets to distinguish between instructions and data.\n
- Capability inventory: The skill uses subprocess calls (mergekit-yaml), file system operations (saving models), and network access (pushing to HuggingFace Hub).\n
- Sanitization: The provided snippets do not include sanitization or validation of model outputs before processing.\n- [EXTERNAL_DOWNLOADS]: The skill references and downloads tools from established technology organizations and repositories.\n
- Evidence: Fetches the mergekit library from Arcee AI's official GitHub repository.\n
- Evidence: Downloads the FastChat framework from the LMSYS organization for model evaluation.\n- [COMMAND_EXECUTION]: The skill instructs the execution of shell commands and Python scripts for model processing and quantization.\n
- Evidence: Uses mergekit-yaml to perform model merging operations.\n
- Evidence: Includes procedures for model quantization using GGUF and GPTQ formats via conversion scripts.
Audit Metadata