The Agent Skills Directory

[DYNAMIC_EXECUTION]: The example provided for creating a Vectorized UDF utilizes pickle.load() to deserialize a model file at runtime.
Evidence (SKILL.md): model = pickle.load(open(sys.path[0] + "/model.pkl", "rb"))
The pickle module is well-known for being vulnerable to arbitrary code execution during deserialization. While the provided example references a local file (model.pkl), this implementation pattern is considered a security risk if the provenance of the file cannot be strictly verified or if the execution environment allows for file tampering.
[INDIRECT_PROMPT_INJECTION]: The skill describes a data transformation workflow that ingests data from external Snowflake tables, creating a surface for indirect prompt injection if the processed data contains instructions intended for an agent.
Ingestion points: The skill reads data from Snowflake tables using session.table("customers") and dbt source/reference macros like {{ ref('stg_orders') }}.
Boundary markers: The skill does not describe or implement delimiters, tagging, or specific instructions to ignore potential commands embedded within the retrieved data.
Capability inventory: The skill utilizes powerful data manipulation capabilities, including row-wise UDFs, vectorized pandas UDFs, and stored procedures that can write data back to the database via save_as_table.
Sanitization: There is no evidence of sanitization, filtering, or validation of the data content to prevent embedded instructions from influencing downstream agent behavior or logic execution.

snowflake-snowpark-dbt