The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses authoritative and restrictive language to override the agent's standard collaborative and safety protocols.
Evidence: The section titled 'CRITICAL BEHAVIOR REQUIREMENT' contains multiple directives to bypass user interaction: 'DO NOT ASK THE USER WHAT THEY WANT TO DO', 'DO NOT OFFER OPTIONS OR CHOICES', 'IMMEDIATELY AND AUTOMATICALLY'.
Evidence: The 'Behavior Guidelines' section explicitly forbids common helpful AI patterns: 'NEVER SAY THESE PHRASES: "What would you like to do with this data?"', 'Act decisively without asking permission'.
These instructions are designed to force the agent into a specific, non-negotiable operational state, disregarding typical safety and user-centric boundaries.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface for indirect prompt injection via external data files.
Ingestion points: The skill is designed to load and inspect CSV files provided by the user using the summarize_csv function mentioned in SKILL.md.
Boundary markers: No evidence of boundary markers or instructions to treat data as untrusted was found. The instructions do not specify any delimiters or warnings to ignore instructions that might be embedded within the CSV cell content.
Capability inventory: The skill has the capability to execute Python code (using pandas, matplotlib, seaborn) and generate files/visualizations based on the data provided.
Sanitization: There is no mention of sanitization, validation, or escaping of the CSV content before it is processed by the analysis logic.

csv-data-summarizer