The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The model implementation in references/custom-pyfunc.md utilizes pickle.load() to deserialize artifacts during the model's context loading phase. This is an unsafe deserialization pattern that could allow for arbitrary code execution if model artifacts were compromised or replaced.
[PROMPT_INJECTION]: The GenAI agent implementation provided in references/genai-agents.md creates a surface for indirect prompt injection.
Ingestion points: Data is ingested from user messages and external retrieval tools like VectorSearchRetrieverTool.
Boundary markers: The provided template lacks explicit delimiters or instructions to the model to ignore potential commands embedded within retrieved documents.
Capability inventory: The agent has the capability to execute tools via Unity Catalog functions.
Sanitization: There is no specific logic included in the templates to filter or sanitize incoming external content before it reaches the language model.
[COMMAND_EXECUTION]: The skill uses the databricks CLI and dbutils functions to manage workspace files and submit serverless jobs. While these are standard administrative actions for the platform, they represent a broad set of execution capabilities managed by the agent instructions.

databricks-ml-training