The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The file references/advanced_features.md contains an experimental code sample that uses npx -y @philschmid/weather-mcp to download and execute code from a third-party npm package at runtime without user confirmation or review.
[PROMPT_INJECTION]: The skill describes tools for ingesting untrusted external data, such as UrlContext in references/structured_and_tools.md and YouTube processing in references/text_and_multimodal.md. These ingestion points lack sanitization or boundary markers to prevent the model from following embedded malicious instructions.
[COMMAND_EXECUTION]: The documentation for ToolCodeExecution in references/structured_and_tools.md allows the model to generate and execute Python code, which could be exploited if the model is influenced by malicious external data.
[PROMPT_INJECTION]: In references/safety.md, a demonstration for safety configuration includes a system instruction to "Be as mean as possible," which serves as an example of prompt-based behavioral override.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of multiple SDKs and libraries, including google-genai, @google/genai, mcp, and pydantic, from public package registries.

gemini-api