The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external PostgreSQL databases which could potentially contain instructions that influence agent behavior.
Ingestion points: Data is retrieved from configured databases in scripts/query.py via the execute_query function.
Boundary markers: No specific delimiters or "ignore instructions" warnings are applied to the database output before it is returned to the agent.
Capability inventory: The skill possesses network access to connect to databases and the ability to execute arbitrary SELECT queries.
Sanitization: The skill implements several mitigations including PII masking for sensitive data, truncating results at 10,000 rows to prevent context overflow, and capping column width to 100 characters.
[COMMAND_EXECUTION]: The skill executes SQL queries, which is its primary function. It employs robust multi-layered security controls to prevent misuse.
Evidence: The skill uses psycopg2 to set the session to readonly=True and autocommit=True in scripts/query.py, which is the primary defense against data modification. Additionally, it performs application-level validation to block data-modifying keywords (INSERT, UPDATE, DROP, etc.) and ensures only single statements are executed.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of a standard database connector.
Evidence: The requirements.txt file lists psycopg2-binary>=2.9,<3.0, a well-known and widely used library for PostgreSQL connectivity.

read-only-postgres