The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for the agent to execute shell commands, including running Docker containers and local Python validation scripts. The YAML frontmatter explicitly permits Bash tool usage for these operations.
[PRIVILEGE_ESCALATION]: The recommended docker run command utilizes --ipc=host and --network=host flags, which reduce container isolation to allow shared memory access and host networking, typical for high-performance GPU tasks. Additionally, the documentation advises against specifying a non-root user ID (--user $(id -u):$(id -g)) because specific TAO toolkit images require the UID to be present in the container's /etc/passwd to function correctly.
[EXTERNAL_DOWNLOADS]: The skill fetches a Docker container image from nvcr.io/nvidia/tao/tao-toolkit, which is the official NVIDIA Container Registry. These downloads are associated with the official vendor and the primary purpose of the skill.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied Parquet embedding files and YAML configuration specs. An attack surface exists if these inputs contain malicious content.
Ingestion points: YAML configuration file (--spec) and Parquet data files (source_path, target_path).
Boundary markers: Not explicitly defined in the prompt interpolation, but the skill relies on rigid tool-based execution.
Capability inventory: The skill can execute shell commands via Docker and write to the local filesystem (output directory).
Sanitization: The included script scripts/verify_unique_neighbor_matching_spec.py performs schema and value validation on the YAML configuration using yaml.safe_load() and strict type checking before execution.

tao-mine-od-images