The Agent Skills Directory

Indirect Prompt Injection: The skill ingests user-provided text (topics) and images (style donors) which are then used to generate prompts for media models and potentially influence script generation. While the instructions include a safety policy to treat this input as data rather than instructions, this remains a potential surface for indirect prompt injection.
Command Execution: The skill utilizes a sandbox_exec tool to perform operations such as file assembly, audio measurement, and network uploads. These commands are executed in a sandboxed environment using scripts located at paths defined by environment variables ($HF_WORKFLOWS). This is a necessary part of the media assembly process but involves the execution of system-level commands.
External Data Interaction: The skill performs network operations via curl to download assets from platform domains and upload the final video to provided URLs. These interactions are directed at infrastructure associated with the service provider (higgsfield.ai) but represent the skill's ability to communicate with external endpoints.

faceless-channel