The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill retrieves and executes installation scripts from official developer domains and repositories associated with the platform vendor.
Downloads a setup script from dev.meta.ai (official developer portal).
Fetches an installation script from the facebookincubator/muse-gadget-sdk GitHub repository.
[COMMAND_EXECUTION]: The skill configures a Linux-based agent environment with tools for remote system management and shell interaction.
Implements the system.run command, allowing the agent to execute shell commands on the linked device.
Includes file.read and file.write commands for remote file system access.
Mitigation: The documentation explicitly informs users that the agent inherits the host account's permissions and recommends installing the service under a non-privileged user account to minimize potential risk.
[INDIRECT_PROMPT_INJECTION]: The skill contains an attack surface for indirect prompt injection by processing external user-supplied content.
Ingestion points: The agent is instructed to fetch and process 'community skills' (Markdown-based instruction files) from a public GitHub repository to extend gadget functionality.
Capability inventory: The agent possesses high-privilege capabilities including shell execution and file system manipulation on the target hardware.
Mitigation: Although the skill lacks formal boundary markers for external data, it provides clear warnings that community integrations are unofficial and should not be used for safety-critical applications.

muse-gadget-setup