The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the snowflake_sql_execute tool to perform sensitive operations, specifically executing GRANT statements to modify database access controls. This functionality requires the user to hold high-privilege roles such as SYSADMIN or ACCOUNTADMIN and involves direct modification of the security environment.
[INDIRECT_PROMPT_INJECTION]: The skill extracts dependency information from Cortex Agent specifications, including free-text instructions. This creates a vulnerability surface where maliciously crafted instructions could lead the agent to audit or grant permissions on unintended database objects.
Ingestion points: Object FQNs are parsed from agent specifications retrieved via the semantic_studio tool (specifically the cortex_agent_read action).
Boundary markers: No specific delimiters or safety warnings are implemented to distinguish agent-provided instructions from validated object names.
Capability inventory: The skill possesses the capability to run SHOW GRANTS and GRANT commands across the account via snowflake_sql_execute.
Sanitization: The instructions do not specify any validation or sanitization logic for the FQNs extracted from the agent's instructions.
[METADATA_POISONING]: There is a discrepancy in the author metadata; the README.md identifies the author as 'Martin Fleischmann', while SKILL.md and the LICENSE file identify the author as 'Martin Seifert'. While likely a simple error, such inconsistencies can be used to obscure ownership or bypass reputation checks.

check-agent-access