The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches source code and project files from a third-party GitHub repository (github.com/krusemediallc/arcads-claude-code.git) that is not associated with the skill author or a trusted organization.
[REMOTE_CODE_EXECUTION]: Instructs the user to run a local setup script (./scripts/setup.sh) and install dependencies from a requirements file (shared/skills/meta-ad-builder/scripts/requirements.txt) provided by the untrusted external repository. This pattern provides a direct path for the execution of unverified third-party code.
[COMMAND_EXECUTION]: Uses subprocess.run to execute system-level commands including ffmpeg, npx (hyperframes), and whisper. These calls use parameters derived from user-controlled data (such as file paths and output names), which presents a risk of command injection if not rigorously sanitized.
[DATA_EXFILTRATION]: Accesses local image and video files to convert them into Base64 strings for transmission to the external API api.arcads.ai. While this is part of the skill's stated functionality, it establishes a data exfiltration vector for local assets.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to its processing of untrusted external data that influences high-privilege capabilities.
Ingestion points: Ingests user-provided advertising prompts, scripts, and dialogue strings within SKILL.md.
Boundary markers: Absent. There are no delimiters or instructions to the agent to ignore embedded commands in the processed data.
Capability inventory: Spawns subprocesses, performs file system writes, and executes network operations.
Sanitization: Absent. The skill lacks evidence of input validation or escaping before data is passed to subprocesses or API payloads.

arcads-claude-code