The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill documentation in SKILL.md, README.md, and CLAUDE.md explicitly directs the installation of the Higgsfield CLI using the command curl -fsSL https://raw.githubusercontent.com/higgsfield-ai/cli/main/install.sh | sh. This pattern allows for the execution of unverified remote scripts with the privileges of the local user.
[COMMAND_EXECUTION]: Python scripts such as scripts/compile-dna.py and scripts/compose-video.py make frequent use of the subprocess module to execute system binaries including ffmpeg, ffprobe, npx, and bash. These commands process data derived from external storyboard files and user-provided assets, creating a path for command injection if inputs are maliciously crafted.
[PROMPT_INJECTION]: The skill contains a vulnerability surface for indirect prompt injection:
Ingestion points: Untrusted data from storyboard Markdown files and raw asset descriptions are read in scripts/compile-dna.py and scripts/compose-video.py.
Boundary markers: The scripts lack boundary markers (such as XML tags) or protective instructions when interpolating this untrusted text into prompts for the Gemini model.
Capability inventory: The skill possesses capabilities to execute shell commands (subprocess), perform network requests (urllib), and manage local files.
Sanitization: No escaping or validation logic is applied to external file content before it is processed by the LLM.
[EXTERNAL_DOWNLOADS]: The script scripts/compose-video.py utilizes urllib.request.urlretrieve to download video files from remote URLs during the generation process.

content-engine