The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes instructions to execute shell commands with elevated privileges using sudo. These are used to create directories, set Access Control Lists (ACLs), and perform file deletions.
Evidence: references/calibration-workflow.md uses sudo setfacl to grant container UIDs access to project directories. references/teardown.md includes a 'Nuke option' using sudo rm -rf to delete bind-mounted application state.
[DATA_EXPOSURE]: The skill manages authentication credentials and accesses sensitive configuration paths on the host system.
Evidence: The skill reads ~/.ngc/config to source API keys and instructs the user to store the NGC_CLI_API_KEY in plain text within .env files for Docker Compose.
[EXTERNAL_DOWNLOADS]: The skill downloads large binary assets and model weights from external repositories.
Evidence: Fetches application data from NVIDIA's NGC registry and downloads the VGGT model from HuggingFace (facebook/VGGT-1B-Commercial). These resources are provided by trusted or well-known organizations.
[INDIRECT_PROMPT_INJECTION]: The skill defines an attack surface by ingesting and processing untrusted external data files which influence downstream agent actions.
Ingestion points: Processes calibration.json, alignment_data.json, and local .mp4 video files.
Boundary markers: No explicit delimiters or warnings are present in the processing logic to ignore embedded instructions in data.
Capability inventory: The skill possesses significant capabilities including docker compose, sudo, and recursive file system deletions.
Sanitization: Uses jq for structured JSON parsing but lacks specific sanitization for content that could influence the LLM logic.

vss-deploy-detection-tracking-3d