The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill takes model names, hub identifiers, and device configurations from user prompts and interpolates them into JSON payloads and shell commands for a local microservice.
Ingestion points: Model details and configuration parameters extracted from the user's prompt (SKILL.md, plugins-guide.md).
Boundary markers: Absent. The skill does not use delimiters or instructions to ignore potential commands embedded in the user's model description.
Capability inventory: Performs network requests via curl to localhost:8200, writes models to the local filesystem, and executes a vendor shell script (SKILL.md).
Sanitization: Absent. User-provided strings are directly used in command-line arguments and JSON fields.
[EXTERNAL_DOWNLOADS]: The skill instructs users to download and execute a shell script from the vendor's official GitHub repository.
Evidence: curl -sSLO https://raw.githubusercontent.com/open-edge-platform/edge-ai-libraries/main/microservices/model-download/scripts/get_model.sh followed by source ./get_model.sh (SKILL.md).
Context: The resource is hosted by the skill's author (open-edge-platform) and is a part of the intended functionality.
[COMMAND_EXECUTION]: The skill generates shell commands for setting environment variables, starting Docker services, and polling job status.
Evidence: Use of source scripts/run_service.sh, docker compose, and watch commands (SKILL.md, service-setup.md).
Note on Automated Scans: Automated scanners flagged piped execution to python3. Manual review confirms these commands use python3 -m json.tool for formatting or python3 -c for parsing specific JSON fields, which are benign operations and do not constitute remote code execution.

model-download-user