The Agent Skills Directory

[DATA_EXFILTRATION]: The skill instructs the agent to perform file discovery in sensitive user directories including ~/Downloads, ~/Desktop, and ~/Pictures using the find command. It also utilizes curl -X PUT --data-binary to upload local files to external URLs provided by the MCP tools. This pattern creates a risk where non-media files could be targeted for exfiltration if the agent is misdirected.
[COMMAND_EXECUTION]: The skill relies on shell commands to manage files and network operations. It uses curl to transfer data and the open command to automatically launch applications for viewing downloaded assets on the user's system.
[EXTERNAL_DOWNLOADS]: The skill downloads media assets from external URLs generated during the processing workflow.
[INDIRECT_PROMPT_INJECTION]: The skill dynamically ingests and acts upon tool descriptions fetched from a connected MCP catalog at runtime, creating an attack surface for instructions embedded in external data.
Ingestion points: Live MCP tool catalog and tool descriptions fetched in Step 2.
Boundary markers: Absent; there are no delimiters to separate data from instructions.
Capability inventory: curl (upload/download), open (file execution/viewing), find (file discovery), and invocation of various MCP tools.
Sanitization: None; the skill does not validate or sanitize the tool descriptions before using them to guide its logic.

media-router