The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses an indirect prompt injection attack surface because it ingests untrusted data from external sources and has capabilities that could be influenced by that data.
Ingestion points: User-provided web URLs, audio/video URLs, and local text/media files as defined in Step 1 of SKILL.md.
Boundary markers: The instructions lack explicit security boundaries or 'ignore' directives to prevent instructions embedded within the ingested material from overriding the agent's system-level constraints.
Capability inventory: The skill can execute shell commands via the listenhub CLI (for image generation and text-to-speech), perform file writes to the current working directory, and make network requests to the vendor's API at api.marswave.ai.
Sanitization: External content is extracted and processed into the LLM context after being parsed as JSON string data, but no specific security sanitization for prompt injection is performed.
[EXTERNAL_DOWNLOADS]: The skill downloads content from arbitrary user-provided URLs to process media locally.
Evidence: Uses curl -L -o to download audio and video files for transcription using the coli tool.
[COMMAND_EXECUTION]: The skill relies heavily on shell command execution to perform its workflow.
Evidence: Orchestrates multiple CLI tools including listenhub for media generation, coli for transcription, jq for data parsing, and standard utilities like base64, mkdir, and curl for file and network operations.

creator