The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to execute an installation script located at ./extensions/banana/install.sh. This script is not included in the skill's file list, creating a risk where the agent may execute unverified code that could download additional payloads from remote sources.
[COMMAND_EXECUTION]: The skill refers to several executable scripts and utilities that are not provided, such as cost_tracker.py and presets.py. Execution of these non-distributed scripts by the agent poses a risk of unauthorized command execution if the environment contains malicious versions of these files.
[EXTERNAL_DOWNLOADS]: The skill is built to interact with the @ycse/nanobanana-mcp server. The source code for this server is hosted on a GitHub repository (github.com/YCSE/nanobanana-mcp) which is not categorized as a trusted source.
[DYNAMIC_EXECUTION]: The skill generates and executes complex shell commands for image processing using magick (ImageMagick) and ffmpeg. While these are standard tools, the commands are constructed dynamically based on user input, which can lead to command injection if not properly handled.
[INDIRECT_PROMPT_INJECTION]: The skill lacks sanitization for user-provided descriptions, which are interpolated directly into prompts for the Gemini image generation API.
Ingestion points: User-provided <description> in SKILL.md.
Boundary markers: None identified; user input is embedded in narrative prompt structures.
Capability inventory: Uses gemini_generate_image via MCP server.
Sanitization: No evidence of input validation or escaping for the prompt components.

seo-image-gen