The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides instructions to download and execute a shell script from a remote URL directly in the terminal using the curl | sh pattern.
[EXTERNAL_DOWNLOADS]: The installation and update processes for the belt CLI involve fetching binary files and manifests from dist.inference.sh.
[DATA_EXFILTRATION]: The documentation describes a feature where providing a local file path (e.g., ~/.ssh/id_rsa or local media) as a model input causes the CLI to automatically upload that file to the platform's cloud servers.
[COMMAND_EXECUTION]: The skill relies on the execution of various shell commands via the belt CLI tool to perform its primary functions.
[COMMAND_EXECUTION]: Instructions for setting up shell completions involve writing configuration data to system-protected directories like /etc/bash_completion.d/, which typically requires root privileges.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface because it ingests untrusted data from the external app store and model results (Ingestion points: app-discovery.md, running-apps.md). It lacks explicit boundary markers to separate this data from instructions. The agent has the capability to execute shell commands (Capability inventory: SKILL.md), and no sanitization of the remote content is specified.

agent-tools