The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill directs the agent to retrieve installation instructions from a remote GitHub repository (github.com/inference-sh/skills) and identifies multiple external skill dependencies.
[REMOTE_CODE_EXECUTION]: The instructions command the agent to install external code using npx skills add for several third-party extensions, including belt-sh/cli and various components from inference-sh/skills. It also facilitates the execution of remote applications via the belt app run interface.
[COMMAND_EXECUTION]: The skill functionality is built around shell command execution using the belt CLI tool for managing sessions (belt login) and running inference tasks.
[DATA_EXFILTRATION]: User input provided to the skill (via prompt or text fields) is sent to the inference.sh remote platform. This creates a potential for data exfiltration if the agent processes sensitive information through these voice synthesis tools.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes untrusted text data for synthesis.
Ingestion points: Text fields within the JSON payloads for the belt app run command in SKILL.md.
Boundary markers: Inputs are encapsulated within JSON structures, which provides a basic boundary but does not prevent instruction following by the underlying model.
Capability inventory: The skill has access to shell execution via the belt tool.
Sanitization: There is no evidence of input sanitization or safety filtering applied to the text before transmission.

ai-voice-cloning