The Agent Skills Directory

[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill fetches the auto-magic-calib source code from the official NVIDIA-AI-IOT GitHub repository and downloads the VGGT-1B-Commercial model from HuggingFace. These external resources originate from trusted and well-known organizations. Additionally, the skill installs the huggingface_hub Python package via pip to facilitate model management.
[PRIVILEGE_ESCALATION]: The instructions involve several operations requiring elevated privileges via sudo, including modifying user group memberships for Docker access, installing system packages (python3-venv, python3-pip), and recursively changing directory ownership (chown) for projects and models. The skill explicitly directs the agent to obtain user confirmation before performing these privileged file system changes.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill facilitates authentication with the NVIDIA container registry and HuggingFace. It implements secure credential handling by passing NGC API keys and HuggingFace tokens through standard input and environment variables, effectively preventing sensitive information from being exposed in shell history or process monitoring tools.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources, creating an attack surface for indirect injection.
Ingestion points: Microservice readiness responses via curl, container registry manifests via docker manifest inspect, and configuration files from the cloned repository.
Boundary markers: The instructions do not define specific boundary markers to isolate untrusted external data.
Capability inventory: The skill can execute subprocesses via docker, git, and python, and perform privileged write operations via sudo chown.
Sanitization: Responses from the microservice endpoint are sanitized by using grep to check for specific success codes, rather than executing the payload directly.

amc-setup-calibration-stack