The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/db_explorer.py facilitates the execution of SQL commands. It mitigates security risks by enforcing read-only access (using ?mode=ro for SQLite and SET SESSION READ ONLY for PostgreSQL), utilizing a regex-based validator (_validate_sql) that permits only SELECT/WITH/EXPLAIN/PRAGMA/SHOW operations, and blocking multi-statement queries through semicolon detection.
[PROMPT_INJECTION]: The skill processes data retrieved from external databases, creating an indirect prompt injection surface if the database content contains instructions intended to manipulate the agent.
Ingestion points: Database content and metadata read via preview, describe_table, and list_tables in scripts/db_explorer.py.
Boundary markers: Not present; the tool returns raw database data which the agent interprets as part of its context.
Capability inventory: The skill allows querying and schema exploration but is restricted by internal SQL validation and database-level read-only permissions.
Sanitization: The skill uses identifier quoting and keyword filtering for query safety, but does not sanitize the values retrieved from the database to prevent downstream prompt injection.
[DATA_EXFILTRATION]: By design, the skill exposes database schema and record content to the agent, which constitutes the movement of potentially sensitive information from the database into the agent's prompt context.
[SAFE]: The skill uses psycopg2-binary, a widely recognized and trusted package for PostgreSQL connectivity. No hidden network communication or suspicious remote execution patterns were found.

database-inspector