The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/collect_snowflake_evidence.py utilizes subprocess.run to execute the Snowflake CLI (snow). This is a functional requirement for gathering authorization evidence. The implementation is secured by:
Using list-based command arguments to prevent shell injection.
Validating all user-supplied identifiers (roles, users, databases) against strict regular expressions (e.g., IDENTIFIER_RE).
[DYNAMIC_EXECUTION]: The orchestrator script scripts/analyze_access_evidence.py uses importlib to dynamically load and execute the core analyzer logic in scripts/analyze_access.py. This is a modular design pattern for local scripts and does not involve loading remote or untrusted code.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process Snowflake authorization metadata which could theoretically contain malicious payloads. This risk is mitigated by a multi-layered evidence chain:
Ingestion points: Snowflake authorization metadata is ingested via snowflake-access-bundle.json into scripts/analyze_access.py.
Boundary markers: The skill instructions and documentation explicitly define the schema for 'sanitized' metadata and evidence receipts.
Capability inventory: The skill possesses Bash (python3) for execution and Write capabilities to generate reports and temporary SQL files for the Snowflake CLI.
Sanitization: The reject_secrets function in analyze_access.py uses regex patterns to identify and block the ingestion of credentials (passwords, tokens, private keys). Additionally, sanitize_output_tree in collect_snowflake_evidence.py recursively redacts sensitive fields before they are written to receipts.

snowflake-access-guardian