The Agent Skills Directory

[SAFE]: The skill implements a robust security model for Snowflake authentication analysis, emphasizing read-only operations and local data processing.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill contains extensive defensive code to detect and reject credentials in input data across multiple scripts (analyze_auth.py, collect_snowflake_evidence.py). It utilizes pseudonymization (SHA-256) for identity values and has no network operations that could facilitate exfiltration.
[COMMAND_EXECUTION]: The collect_snowflake_evidence.py script uses subprocess.run to invoke the Snowflake CLI (snow). Inputs, such as connection profiles, are strictly validated against regular expressions to prevent command injection, and the command structure is static.
[INDIRECT_PROMPT_INJECTION]: The skill processes Snowflake evidence files (JSON), which represents a potential attack surface. However, the skill mitigates this by using deterministic Python scripts for all data processing and implementing deep sanitization and redaction logic for all output structures, preventing malicious data from influencing the agent's context.
[DYNAMIC_EXECUTION]: The use of importlib.util in analyze_auth_evidence.py for loading internal scripts is performed using static local paths for code modularity and does not allow for arbitrary code execution.
[PROMPT_INJECTION]: The static detections in eval-spec.yaml are false positives; the identified strings are test cases designed to evaluate the skill's ability to resist injection, rather than instructions to the agent during runtime.

snowflake-strong-auth-migration-pilot