The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill recommends installing the infsh CLI using a piped shell execution pattern: curl -fsSL https://cli.inference.sh | sh. This pattern is highly dangerous as it downloads and executes code from a remote server directly in the terminal without any prior verification or integrity checks.
[DATA_EXFILTRATION]: The CLI tool explicitly supports uploading local files to the inference.sh platform. The documentation in references/running-apps.md provides examples for uploading files from absolute and relative paths. If an attacker provides a path to sensitive files (e.g., ~/.ssh/id_rsa, ~/.aws/credentials, or .env files), the agent could be tricked into exfiltrating this data to the remote service.
[INDIRECT_PROMPT_INJECTION]: The skill handles untrusted data without sufficient security boundaries:
Ingestion points: User-provided prompts are interpolated directly into the JSON --input argument of the infsh app run command.
Boundary markers: No delimiters or instructions are provided to the agent to distinguish between the prompt data and the commands, nor is the agent instructed to ignore instructions embedded within the user data.
Capability inventory: The skill utilizes the terminal tool to execute shell commands and has the capability to read local files for upload.
Sanitization: The instructions do not specify any sanitization, escaping, or validation of the user prompt before it is included in the shell command execution.
[COMMAND_EXECUTION]: The skill makes extensive use of the terminal tool to run infsh commands. It includes instructions for writing shell completions to system-level directories (e.g., /etc/bash_completion.d/infsh), which modifies system configuration and typically requires elevated privileges.

inference-sh-cli