The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from multiple file formats (PDF, DOCX, CSV, images), which constitutes a potential attack surface. However, it incorporates safe handling practices: \n
Ingestion points: Data is ingested from user-provided files such as target_data.csv, sales_2023.csv, and feedback.pdf using Pandas and specialized platform tools. \n
Boundary markers: The skill mandates a 'Safety exploration' (安全探查) phase to read only metadata and limited samples (n=5), effectively limiting the ingestion of potentially malicious instructions into the prompt context. \n
Capability inventory: The skill generates and executes Python code for data processing and visualization, and produces reports in Markdown, PDF, and DOCX formats. \n
Sanitization: The guidelines include mandatory PII de-identification (脱敏) for fields like IDs and phone numbers and rigorous data quality validation. \n- [SAFE]: The skill uses standard data science libraries (Pandas, Matplotlib, Plotly) and refers to official toolsets without any evidence of credential theft, remote code execution from unknown sources, or persistence mechanisms.

data-analysis