The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/mirror_image.py script executes system commands via the subprocess module to interact with docker and the aws CLI. Analysis confirms that commands are passed as argument lists (e.g., subprocess.run([docker, 'pull', public_uri])) rather than shell strings, which effectively prevents shell injection vulnerabilities. The script also follows security best practices by passing sensitive tokens to docker login via stdin rather than as command-line arguments.
[EXTERNAL_DOWNLOADS]: The skill references several external resources, including the AWS Deep Learning Containers catalog (aws.github.io), GitHub API for repository metadata, and Hugging Face model configurations. All referenced domains are trusted organizations or well-known services. The mirror_image.py script performs image pulls from public.ecr.aws and ghcr.io, which are standard, trusted registries for the intended infrastructure tasks.
[INDIRECT_PROMPT_INJECTION]: The skill defines a data ingestion surface by instructing the agent to fetch and parse config.json files from Hugging Face model repositories (huggingface.co) to determine the appropriate serving architecture.
Ingestion points: Model configuration files fetched from huggingface.co via curl.
Boundary markers: None explicitly defined in the prompt interpolation.
Capability inventory: The agent has access to mirror_image.py, which can perform Docker and ECR operations.
Sanitization: The agent is directed to specifically extract the architectures key from the JSON structure, reducing the risk of accidental instruction following from other metadata fields. Given the narrow scope of the data being read and the trusted nature of the source platform, the risk is negligible.

hf-cloud-serving-image-selection