The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs users to install its core command-line interface by piping a remote shell script directly into the interpreter (curl -fsSL https://raw.githubusercontent.com/higgsfield-ai/cli/main/install.sh | sh). This pattern is a critical security risk as it executes remote code without allowing the user to verify the script's content beforehand.
[EXTERNAL_DOWNLOADS]: The skill and its references guide the user to download multiple external resources at runtime:
Fetches 3D game assets and animations via git clone from an unverified third-party repository (KayKit-Game-Assets).
Downloads a portable Blender binary and multiple system dependencies (.deb packages) from external sources to handle 3D processing in a sandbox environment.
The compose_cover.py script automatically fetches font files and branded wordmark assets from static.higgsfield.ai during execution.
[COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool to orchestrate a complex build pipeline. This includes running system-level utilities like ffmpeg and Blender, as well as a large suite of included Python scripts designed to manipulate 3D models and image textures. The broad scope of these executions increases the risk of malicious command manipulation.
[PROMPT_INJECTION]: The skill possesses a significant attack surface for indirect prompt injection. It is designed to ingest detailed user-supplied 'design briefs' and 'intake' data, which are then used to drive file-writing and command-execution logic. The instructions do not include robust boundary markers or sanitization protocols to prevent adversarial content within these briefs from overriding the agent's safety guidelines or core mission parameters.
Ingestion points: Design briefs and user intake responses (processed in references/website-flow.md and references/app-flow.md).
Boundary markers: Absent; the skill directly interpolates descriptive text from briefs into prompts for asset generation and code construction.
Capability inventory: Full filesystem access, network requests (curl/fetch), and arbitrary command execution (bash/python).
Sanitization: No explicit filtering or validation of user-provided copy before it is used in system-level operations.

higgsfield-websites