The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection by ingesting untrusted data from external URLs and web search results, which could potentially contain adversarial content intended to influence the agent's behavior.
Ingestion points: The skill accepts image_urls for image-to-image tasks and supports an enable_web_search feature in Nano Banana models (SKILL.md).
Capability inventory: The skill executes the runcomfy CLI via Bash and writes generated image files to the local file system (SKILL.md).
Boundary markers: There are no structural delimiters enforced on the untrusted data, though the documentation provides instructions for the agent to suspect reference assets if output diverges from the prompt.
Sanitization: Prompt content is handled as a JSON string to prevent shell injection, but the semantic content of external inputs remains a vector for model steering.
[EXTERNAL_DOWNLOADS]: The skill references the installation of the @runcomfy/cli package from the NPM registry and fetches generated assets from the vendor's official domains (*.runcomfy.net and *.runcomfy.com).
[COMMAND_EXECUTION]: The skill utilizes the runcomfy CLI tool to perform its primary functions, which is constrained by the allowed-tools restriction to prevent the execution of unrelated shell commands.

ai-image-generation