The Agent Skills Directory

Remote Code Execution (CRITICAL): The Quick Start section includes the command curl -fsSL https://cli.inference.sh | sh. Piped shell execution of remote scripts is an extremely dangerous pattern as it allows for immediate, unverified code execution with the privileges of the current user.
External Downloads (HIGH): The skill installs a CLI and downloads 'apps' (e.g., falai/dia-tts) from inference.sh, which is not a trusted source. It also uses npx skills add to fetch external skill definitions from the same provider.
Command Execution (MEDIUM): The allowed-tools configuration allows the Bash tool to run any infsh command (infsh *). This provides a broad attack surface, allowing the agent to execute any functionality within the untrusted CLI.
Indirect Prompt Injection (LOW): The skill processes untrusted user-provided text data for audio generation, which could be exploited to influence downstream agent actions if the audio output or logs are processed further.
Ingestion points: The prompt field in the infsh app run commands.
Boundary markers: None. No delimiters or instructions to ignore embedded commands are present.
Capability inventory: Subprocess execution via Bash(infsh *).
Sanitization: None detected in the documentation or provided examples.

dialogue-audio