The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill allows the agent to download and install command-line tools from arbitrary GitHub repositories using the advai cli install <github-url> command. While GitHub is a well-known hosting service, the code within third-party repositories is not inherently verified by the skill.
[REMOTE_CODE_EXECUTION]: By design, the skill executes software retrieved from remote GitHub repositories. Once installed, these tools are run via advai cli <name> ..., effectively executing code from an external source on the local system.
[COMMAND_EXECUTION]: The skill invokes shell commands through the advai and opencli tools. It specifically executes subcommands and third-party binaries, providing a pathway for arbitrary command execution depending on the installed tools.
[PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection because it ingests and processes data from external CLI tools (e.g., via advai cli info <name> or tool output). Malicious instructions embedded in a third-party tool's documentation or output could attempt to influence agent behavior.
Ingestion points: Output and metadata from external CLIs managed by the advai cli toolchain.
Boundary markers: No explicit boundary markers or instructions to ignore embedded content are defined in the skill logic.
Capability inventory: The skill possesses the capability to execute sub-processes, install new software, and perform network requests (via the install command).
Sanitization: There is no evidence of sanitization or validation performed on the data returned by the external CLI tools before it is processed by the agent.

advai-cli