The Agent Skills Directory

[COMMAND_EXECUTION]: The skill facilitates the execution of shell commands on remote cloud instances through its primary CLI tool, SkyPilot. This includes setup commands for installing dependencies and run commands for executing training or inference workloads. This behavior is the intended primary purpose of the skill.
[EXTERNAL_DOWNLOADS]: The instructions guide the user to install the SkyPilot package from the official Python Package Index (PyPI) and fetch Helm charts from the vendor's official repository at https://helm.skypilot.co. These sources are official and qualify as safe under the trusted vendor rules.
[DATA_EXFILTRATION]: Instructions include commands for configuring cloud credentials (e.g., aws configure, gcloud auth login, az login) and accessing SSH keys (~/.ssh/sky-ssh-config). This behavior is necessary for the tool's intended purpose of managing cloud infrastructure and uses standard provider utilities.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it processes user-provided configuration files and synchronizes local content to remote clusters. * Ingestion points: task.yaml configuration files, workdir local or Git repository content, and cloud storage buckets mounted via file_mounts. * Boundary markers: The skill employs a secrets field in the YAML specification to ensure sensitive environment variables are redacted in logs and dashboards. * Capability inventory: The tool enables full shell access to remote cloud instances, file system modifications on remote nodes, and network operations from those nodes. * Sanitization: User-defined sensitive data in the secrets section is redacted from output logs.

skypilot