The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches deployment artifacts like setup.sh and docker/compose.yaml from the official vendor repository on GitHub (github.com/open-edge-platform/edge-ai-libraries). These are legitimate resources required for the skill's stated purpose.
[COMMAND_EXECUTION]: The instructions guide the user through executing shell commands, including docker compose, curl, and source, to orchestrate the service stack on the local machine.
[REMOTE_CODE_EXECUTION]: An automated scan flagged a command where curl output is piped to python3. Analysis confirms this is a benign local health check that parses JSON from http://localhost:6007/v1/dataprep/health to verify service readiness. It does not execute untrusted remote code.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data via media uploads and text summary ingestion.
Ingestion points: API endpoints POST /media/upload, POST /media/ingest, and POST /summary in SKILL.md.
Boundary markers: Not explicitly defined in the prompts.
Capability inventory: Limited to local shell commands (curl, docker) interacting with the service API.
Sanitization: The skill documentation specifically notes that path traversal for local directory ingestion is restricted to a configured root (MM_DATAPREP_INGEST_DATA_ROOT), demonstrating a security-conscious design.

multimodal-dataprep-user