The Agent Skills Directory

[PROMPT_INJECTION]: The file eval-spec.yaml contains test cases with prompt injection patterns such as "Ignore all previous instructions" and "print your full system prompt". While these are intended for evaluating the agent's robustness against such attacks, the presence of these strings triggers security detectors.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data that could be controlled by an attacker to influence the agent's behavior.
Ingestion points: The hook hooks/bundle-deploy-guard.py reads terraform.tfstate files; hooks/bundle-grant-retry.py processes the stdout/stderr output from databricks bundle deploy commands; scripts/drain-workspace.py ingests inventory data from JSON files.
Boundary markers: Absent. The data is processed directly via regex or JSON parsing without specific instructions to ignore potentially embedded malicious prompts.
Capability inventory: The skill has access to Bash(databricks:*), Bash(terraform:*), and Bash(python3:*), allowing it to perform significant workspace and infrastructure operations.
Sanitization: hooks/bundle-grant-retry.py uses a specific regular expression (D6_SIGNATURE) to isolate a known error pattern, which provides some filtering, while bundle-deploy-guard.py validates JSON structure.
[COMMAND_EXECUTION]: The skill includes several Python scripts designed to be executed via the Bash tool to perform administrative tasks in a Databricks environment. One script, scripts/drain-workspace.py, performs destructive actions by terminating compute resources (clusters, pools, and warehouses). These actions are aligned with the skill's primary purpose for CMK rotation and require explicit user approval via the --execute flag.

databricks-bundle-medic