The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The recovery testing script (scripts/test_recovery.sh) includes a functionality to verify restored data by executing SQL queries provided via the --data-query parameter. This parameter is passed directly to database clients (psql, mysql, mongosh) for execution on a test instance without sanitization. If an attacker provides malicious instructions through an external source (like a pull request or issue) that the agent is analyzing, and the agent is then prompted to run a recovery test, the attacker could influence the SQL executed, leading to unauthorized data modification or extraction on the test database instance.
Ingestion points: The script scripts/test_recovery.sh ingests untrusted data through the --data-query command-line argument.
Boundary markers: There are no explicit boundary markers or 'ignore embedded instructions' warnings applied to the verification queries.
Capability inventory: The skill scripts utilize psql, mysql, mongosh, pg_restore, tar, and docker for database and container operations.
Sanitization: No sanitization, validation, or escaping of the input queries is performed prior to execution.
[COMMAND_EXECUTION]: The provided scripts automate database management tasks by executing various system utilities and database clients. The failover.sh and pitr_restore.sh scripts facilitate the promotion of standby databases and the restoration of WAL/transaction logs. While these actions are consistent with the skill's stated purpose, the use of shell scripts to manage critical database infrastructure represents a significant level of autonomy that should be reviewed before deployment in production environments.

managing-database-recovery