The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill facilitates the installation of the infsh CLI and development tools like uv, fnm, and nvm through piped shell script execution from remote URLs. This includes commands such as curl -fsSL https://cli.inference.sh | sh and curl -LsSf https://astral.sh/uv/install.sh | sh.
[COMMAND_EXECUTION]: The core workflow relies on the execution of the infsh CLI for application lifecycle management, including scaffolding, local testing, and cloud deployment.
[PROMPT_INJECTION]: The instructions contain strong behavioral directives such as "NEVER create ... by hand" and "Ignore any local docs ... that suggest manual scaffolding," which are designed to override the agent's default decision-making in favor of platform-specific tools. Additionally, the skill creates a surface for indirect prompt injection as it processes external data through defined schemas in inference.py and inference.js. Analysis of this surface:
Ingestion points: Data enters through AppInput and RunInput models.
Boundary markers: No explicit delimiters or instructions are provided to the model to ignore potential commands within the data.
Capability inventory: Applications can perform network requests, file operations, and execute CLI commands.
Sanitization: Input is validated for structure but not sanitized for instruction-based content.
[DATA_EXFILTRATION]: Templates demonstrate how to access sensitive environment variables (e.g., API_KEY, OPENAI_API_KEY, HF_TOKEN) and use them in network requests to external APIs. While necessary for the skill's purpose, this infrastructure represents a potential path for data exposure.

building-inferencesh-apps