The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run and subprocess.Popen to interact with the cortex CLI and internal Python scripts. These calls are implemented using argument lists with shell=False, which effectively prevents command injection vulnerabilities.
[CREDENTIALS_UNSAFE]: The skill implements a robust credential protection mechanism. A credential_file_allowlist is used to identify and block prompts containing sensitive paths such as SSH keys, cloud provider credentials, and environment files. This prevents accidental exposure of local secrets to the AI model.
[PROMPT_INJECTION]: The PromptSanitizer module provides multi-layered protection against prompt injection. It uses regex patterns to identify common bypass attempts and employs Unicode normalization (NFKC) to detect obfuscated instructions hidden via homoglyphs or zero-width characters.
[DATA_EXFILTRATION]: Execution is restricted to local processing and interactions with the official Snowflake Cortex CLI. The skill includes PII removal for emails, credit cards, and API keys before prompts are sent for processing.
[EXTERNAL_DOWNLOADS]: All external dependencies, such as the cortex CLI itself, are referenced as prerequisites from official Snowflake documentation. The skill does not perform any unverified remote script execution.
[SAFE]: The skill follows security best practices by implementing structured audit logging, secure cache management with HMAC-SHA256 integrity checks, and granular security envelopes (Read-Only, Read-Write) that restrict the available toolset based on the operation risk.

cortex-code