The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to install official packages from Google (google-generativeai for Python and @google/generative-ai for Node.js) to interact with the Gemini API. These are well-known libraries from a trusted organization.
[DATA_EXFILTRATION]: The skill demonstrates how to configure the API using environment variables (e.g., os.environ['GEMINI_API_KEY']) rather than hardcoding secrets. It includes a specific warning to never hardcode API keys, which is a standard security best practice.
[PROMPT_INJECTION]: As the skill is designed to build agent workflows and process external data (such as images via Vision or text via prompts), it naturally possesses an indirect prompt injection surface. The skill includes examples of 'Agent Loops' and 'Tool Use' (function calling) which could be targeted by malicious instructions embedded in processed data. However, this is a standard capability for LLM integrations and no malicious implementation was found.
Ingestion points: model.generate_content(), chat.send_message(), and PIL.Image.open() (SKILL.md).
Boundary markers: None provided in the examples.
Capability inventory: File system read (vision example), network operations (API calls via SDK), and tool execution (function calling example).
Sanitization: None present in the boilerplate examples. Developers should implement their own validation when deploying these patterns.

gemini-api