The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides numerous templates for ffmpeg and ffprobe commands that involve variable interpolation (e.g., {int}, {duration}). The agent is instructed to execute these commands locally. If the filenames or parameters are not strictly sanitized by the agent platform, this could lead to command injection attacks against the host environment where the agent is running.
[DYNAMIC_EXECUTION]: The instructions require the agent to dynamically construct shell command strings at runtime for video trimming, audio extraction, and frame analysis. This practice, while functional for video editing, introduces risks if the input variables originate from untrusted sources without validation.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection as it processes untrusted "Reference videos" and "Product images".
Ingestion points: The skill accepts external video files and images from the user or remote URLs (e.g., SKILL.md Phase A and B).
Boundary markers: The prompt structure in Phase D lacks explicit delimiters or "ignore" instructions for the product descriptions extracted from these images.
Capability inventory: The skill possesses the capability to execute shell commands (ffmpeg), write files to the local system, and call external generation APIs (arcads_generate_video_omni_flash).
Sanitization: There is no evidence of sanitization or filtering logic for the text extracted from the reference media before it is used in downstream prompts or analysis tools.
[EXTERNAL_DOWNLOADS]: The documentation explicitly mentions the use of third-party video downloading services (e.g., "ssstik") to obtain reference materials. Interacting with these unverified external services to fetch content for local processing introduces risks of malware or malicious payloads embedded in the downloaded files.

viral-clone