The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute arbitrary code (Python and SQL) on remote Databricks clusters using MCP Command Execution tools.
[COMMAND_EXECUTION]: The 'Execution Rules' section explicitly mandates that commands should run automatically without seeking user confirmation, which removes the human-in-the-loop and increases the risk of the agent executing destructive or unauthorized operations if influenced by malicious input.
[REMOTE_CODE_EXECUTION]: The skill facilitates the generation and execution of code on external infrastructure (Databricks), which is a high-privilege capability involving remote environments.
[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection as it processes untrusted data and translates it into executable commands.
Ingestion points: User-provided code snippets and data fetched from Databricks tables during debugging or validation tasks (SKILL.md).
Boundary markers: Absent; there are no instructions provided to separate data from instructions or to treat cluster output as untrusted.
Capability inventory: Arbitrary Python and SQL execution on Databricks clusters via databricks_command and execute_command_with_context (SKILL.md).
Sanitization: Absent; the instructions direct the agent to re-run fixed code based on cluster errors without validation or sanitization of the generated logic.

databricks-testing