The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to install the belt CLI using a 'curl piped to shell' pattern (curl -fsSL https://cli.inference.sh | sh). This method executes remote scripts directly and is an insecure installation practice, though the skill provides manual verification steps as an alternative.\n- [DATA_EXFILTRATION]: The belt CLI features automatic local file uploads when a file path is provided in the input JSON for an app run (e.g., belt app run ... --input '{"image": "/path/to/photo.jpg"}'). This functionality could be exploited to exfiltrate sensitive files like SSH keys (~/.ssh/id_rsa) or environment variables (.env) if the agent is tricked into using those paths.\n- [COMMAND_EXECUTION]: The skill requires the execution of the belt binary via the shell to interact with the platform, providing a direct interface for the agent to perform network and file system operations.\n- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it processes untrusted inputs which are then used to trigger file reading and network transmissions.\n
Ingestion points: JSON input parameters for the belt app run command.\n
Boundary markers: None identified in the instructions or reference files.\n
Capability inventory: Shell command execution, file reading, and remote data uploading.\n
Sanitization: There is no evidence of sanitization or validation for file paths or JSON content provided to the CLI.

infsh-cli