The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data from potentially untrusted local files (CSV, JSON, etc.) and remote sources (S3, HTTPS).
Ingestion points: Files provided via the $0 argument in SKILL.md are ingested by the agent for conversion.
Boundary markers: No delimiters or instructions are used to distinguish data from potential embedded instructions within the files.
Capability inventory: The skill uses Bash to execute find and duckdb, allowing for file system read/write and network operations.
Sanitization: There is no validation or sanitization of the input file content before it is processed by the DuckDB engine.
[DYNAMIC_EXECUTION]: The skill dynamically constructs shell and SQL commands by interpolating user-provided arguments, which can lead to injection.
Shell Injection: In Step 1, the $0 argument is interpolated directly into a find command. If a user provides a filename containing shell metacharacters (e.g., $(command)), it may lead to unintended command execution.
SQL Injection: In Step 2, the input and output paths are placed inside single quotes within a DuckDB SQL string (FROM '<INPUT_PATH>'). If a filename contains a single quote, it could break the SQL literal and allow execution of arbitrary SQL commands within the DuckDB process.

convert-file