The Agent Skills Directory

[COMMAND_EXECUTION]: The utility scripts scripts/inspect.js and scripts/query.js take raw SQL and table names as command-line arguments. These inputs are used to construct and execute database queries (e.g., db.all(sql), db.run(sql), and PRAGMA statements) without internal sanitization or parameterization. This allows for arbitrary database operations if the calling agent does not validate the input.
[DATA_EXFILTRATION]: The skill can be used to open any local file path provided to the db-path argument via path.resolve(). If an agent is directed to a sensitive system file, the inspect.js script may attempt to reveal its structure or content by treating it as a database file, potentially leading to information disclosure.
[PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by ingesting untrusted data from SQLite databases. Data read from tables is returned to the agent without boundary markers or sanitization, potentially allowing malicious content stored within a database to influence the agent's logic.
Ingestion points: process.argv in scripts/inspect.js and scripts/query.js for paths and queries; database records read via db.all() and db.get() methods.
Boundary markers: None implemented in the provided scripts or documentation to distinguish database data from system instructions.
Capability inventory: File system read/write access through the SQLite driver and arbitrary SQL execution.
Sanitization: No validation of SQL syntax or restriction on file paths before execution.

sqlite-client