The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/connect.py contains a critical security vulnerability where it offers to install the Higgsfield CLI by downloading a script from an external GitHub URL (https://raw.githubusercontent.com/higgsfield-ai/cli/main/install.sh) and piping it directly to a shell (sh). This allows for arbitrary code execution from a source that is not recognized as a trusted vendor. Evidence: The PROVIDERS dictionary in scripts/connect.py defines the installation command, and the connect function executes it via subprocess.run(spec["install"], shell=True).
[COMMAND_EXECUTION]: Multiple scripts within the skill utilize the Python subprocess module to execute external binaries and CLI tools. Evidence: scripts/creator.py calls CLI tools for higgsfield, agy, and codex. scripts/media.py orchestrates local media operations via ffmpeg, ffprobe, and rembg. scripts/connect.py executes the gcloud CLI and various other OAuth login sequences.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from user-provided URLs or pasted content to generate ad scripts and media, which can be exploited to influence agent behavior.
Ingestion points: SKILL.md (Step 1) instructs the agent to collect product details and URL content directly from the user for processing.
Boundary markers: There are no explicit boundary markers or warnings used when interpolating external content into model prompts.
Capability inventory: The skill has extensive capabilities including file system access, network operations via urllib.request, and shell command execution via subprocess.
Sanitization: Sanitization is insufficient; while media.py escapes paths for ffmpeg, natural language inputs used in prompts are not sanitized or filtered.

creator-ad