The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The scripts/connect.py script facilitates the installation of the Higgsfield CLI by downloading a shell script from https://raw.githubusercontent.com/higgsfield-ai/cli/main/install.sh and piping it directly to sh. This enables the execution of arbitrary remote code from a source that is not recognized as a trusted organization or well-known service.\n- [EXTERNAL_DOWNLOADS]: The skill attempts to download external binaries and installation scripts. While downloads from trusted sources like Google and OpenAI are handled safely, the Higgsfield download originates from an untrusted third-party repository, posing a supply chain risk.\n- [COMMAND_EXECUTION]: The scripts scripts/creator.py and scripts/media.py make extensive use of subprocess.run to execute system tools such as ffmpeg, ffprobe, and fc-match, as well as various AI provider CLIs. These commands interpolate user-provided data, such as image prompts and file paths, without sufficient validation to prevent command injection.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied branding information and vibe descriptions to generate prompts for downstream AI models, creating a vulnerability to indirect prompt injection.\n
Ingestion points: User inputs for brand name, personality, and audience in SKILL.md.\n
Boundary markers: The generation instructions in references/image-prompting.md lack explicit "ignore embedded instructions" warnings or robust delimiters.\n
Capability inventory: The skill has access to extensive shell execution capabilities and network operations through the creator.py and media.py scripts.\n
Sanitization: Minimal sanitization is performed on user inputs before they are interpolated into generative prompts or shell commands.

creator-brandkit