The Agent Skills Directory

[METADATA_POISONING]: Several scripts in the scripts/ directory contain identical implementation logic that contradicts their individual descriptions. audit_log_generator.py (which claims to generate audit log entries) and audit_table_creator.py (which claims to create database tables) both contain the exact code for a filesystem directory scanner (Analyzer class using rglob('*')). This deceptive metadata masks the actual behavior of the scripts, although the behavior itself (recursive file metadata scanning) is not explicitly malicious.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to have the agent interact with and analyze database audit logs which contain data from various external sources and operations.
Ingestion points: Database logs accessed via audit_log_analyzer.py or through the agent's Read and Grep tools.
Boundary markers: The instructions do not specify the use of delimiters or 'ignore' instructions when the agent processes logs that might contain adversarial text.
Capability inventory: The skill allows execution of Bash(psql:*), Bash(mysql:*), and general filesystem tools (Write, Edit), providing a significant impact surface if the agent is manipulated by log content.
Sanitization: There is no logic provided to sanitize or filter potential control characters or instructions embedded within the database record values being audited.

implementing-database-audit-logging