The Agent Skills Directory

[COMMAND_EXECUTION]: The scheduled_tasks.md file provides a code example using the Node.js child_process module to execute system-level commands, such as pg_dump for database backups. This pattern allows for arbitrary command execution if parameters are not strictly validated.
[REMOTE_CODE_EXECUTION]: The skill describes an 'AI Agent' pattern where an LLM dynamically generates SQL queries that are passed directly to database nodes for execution. This creates a high-risk surface for LLM-driven code execution.
Evidence: Examples in ai_agent_workflow.md and database_operations.md show queries configured as {{$json.sql}}, where the SQL is provided by the AI agent output.
[PROMPT_INJECTION]: The skill defines an architecture for processing untrusted external data through AI agents, which is vulnerable to indirect prompt injection.
Ingestion points: webhook_processing.md (incoming HTTP request bodies) and ai_agent_workflow.md (retrieved data from vector stores/document loaders).
Boundary markers: The provided system prompt templates in ai_agent_workflow.md do not include explicit instruction-isolation delimiters or 'ignore' safeguards for processing external data.
Capability inventory: The workflows have access to SQL execution (database_operations.md), shell command execution (scheduled_tasks.md), and external network requests (http_api_integration.md).
Sanitization: The skill recommends using read-only database users and provides logic for scanning AI-generated SQL for dangerous keywords like 'DROP' or 'DELETE' in ai_agent_workflow.md.

n8n-workflow-patterns