The Agent Skills Directory

[SAFE]: The skill author ('jeremylongshore') is a recognized vendor resource, and the skill's behavior is entirely consistent with its stated purpose of Snowflake cost auditing.
[DATA_EXPOSURE]: The skill demonstrates best-in-class data privacy practices. SQL templates (e.g., scripts/sql/cost.sql) use SHA2 hashing within Snowflake to pseudonymize sensitive identifiers such as usernames, warehouse names, and query tags before the data is exported to the local environment.
[COMMAND_EXECUTION]: The scripts/collect_snowflake_evidence.py script uses subprocess.run to interface with the official Snowflake CLI (snow). This is a legitimate functional requirement, and the script includes a FORBIDDEN_SQL blacklist (blocking commands like DROP, DELETE, and GRANT) and a SAFE_START whitelist (SELECT, SHOW, WITH, DESCRIBE) to ensure only read-only operations are performed.
[CREDENTIALS_UNSAFE]: The skill contains active defensive logic to prevent the ingestion or exposure of secrets. The collector and analyzer scripts use regex-based redaction (e.g., SENSITIVE_KEYS and redact_authorization_values) to strip authentication headers, tokens, and passwords from all evidence and error logs.
[INDIRECT_PROMPT_INJECTION]: While the skill ingests external data from Snowflake, the risk of indirect prompt injection is mitigated by the use of a deterministic Python analyzer (scripts/analyze_cost_evidence.py). The LLM coordinates the workflow, but the actual data processing and report generation are handled by code that does not execute embedded instructions from the input data.
[OBFUSCATION]: The skill uses cryptographic hashing for privacy, but no malicious obfuscation techniques (such as hidden URLs, homoglyphs, or encoded payloads) were detected in the instructions or scripts.
[PRIVILEGE_ESCALATION]: The skill explicitly instructs the agent to avoid requesting ACCOUNTADMIN privileges and to report missing access rather than attempting to escalate permissions.

snowflake-cost-leak-hunter