The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text for prompts and search queries. It explicitly acknowledges that these inputs can come from untrusted channels and provides mandatory instructions for the agent to escape and double-quote variables to prevent shell expansion or injection attacks. This awareness of the attack surface and provided mitigations align with security best practices for prompt interpolation.
[COMMAND_EXECUTION]: The skill utilizes the comfy CLI for various operations, including file system management (creating projects, decomposing workflows) and executing remote jobs. These actions are within the intended scope of the tool's functionality for media generation and workflow orchestration.
[DATA_EXFILTRATION]: While the skill performs network operations such as uploading assets and submitting jobs, these are directed to the vendor's infrastructure (e.g., api.comfy.org) or local endpoints (127.0.0.1) as part of its core functionality. The instructions specifically forbid manual extraction of API keys, emphasizing secure credential management through the CLI's internal auth flow.

comfy