The Agent Skills Directory

[EXTERNAL_DOWNLOADS] (SAFE): The skill instructs the user to install skypilot and related machine learning libraries from standard package registries. These are well-known and legitimate tools.
[COMMAND_EXECUTION] (SAFE): The skill demonstrates how to use SkyPilot to launch clusters and run distributed training jobs. These operations involve standard shell command execution for infrastructure orchestration.
[DATA_EXFILTRATION] (SAFE): The documentation mentions handling cloud credentials and environment variables (e.g., HF_TOKEN, WANDB_API_KEY). However, it does not contain any logic to extract or send these secrets to unauthorized external locations.
[INDIRECT_PROMPT_INJECTION] (LOW): The skill describes features that ingest data from external sources, creating a potential attack surface.
Ingestion points: Git repositories specified in workdir and cloud storage buckets (S3/GCS) specified in file_mounts in both SKILL.md and references/advanced-usage.md.
Boundary markers: Absent; the configurations do not specify markers to distinguish between trusted and untrusted content within these external sources.
Capability inventory: The skill allows for arbitrary shell command execution via the run and setup blocks in YAML configurations.
Sanitization: Absent; there is no mention of validating or sanitizing the content pulled from remote repositories before execution.

skypilot-multi-cloud-orchestration