The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill accesses sensitive Databricks billing and compute system tables. This access is performed via the official Databricks CLI and requires explicit administrative grants, following security best practices. No unauthorized exfiltration to third-party domains was detected.
[COMMAND_EXECUTION]: Shell commands are used to invoke the databricks CLI and jq for data retrieval and processing. These tools are standard for the skill's intended environment and are used to interact solely with the user's configured Databricks workspace.
[DYNAMIC_EXECUTION]: A provided Python script performs deterministic data ranking and report generation. This is a safety measure to ensure arithmetic accuracy and prevent the model from hallucinating financial figures.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from external APIs. Risks are mitigated by the use of structured JSON parsing and deterministic processing logic rather than direct prompt interpolation.
Ingestion points: databricks CLI query results and databricks-workspace-mcp tool outputs.
Boundary markers: Data is passed as JSON directly to the deterministic ranker script.
Capability inventory: Shell command execution (Bash), file writing (Write), and file editing (Edit).
Sanitization: Data is parsed using the standard Python json library and specific numeric fields are extracted and cast in rank-and-report.py, minimizing the risk of instruction execution through malformed data.

databricks-cost-leak-hunter