The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains meta-instructions intended to update the agent's internal knowledge base and override older training data (e.g., 'This skill corrects stale training data'). These are standard for knowledge-delta skills and do not attempt to bypass safety guidelines.
[PROMPT_INJECTION]: The skill enables and documents tools for ingesting untrusted external data, which is a primary vector for indirect prompt injection.
Ingestion points: The skill introduces url_context (fetching live URLs) and file_search (managed RAG) as tools for the agent to process external content (see SKILL.md and references/embeddings-and-rag.md).
Boundary markers: The provided code snippets do not demonstrate the use of specific boundary markers (e.g., XML tags or delimiters) to isolate untrusted content from instructions.
Capability inventory: The skill enables the agent to perform network fetches (URL context), file read/write (File Search API), and UI interactions (Computer Use).
Sanitization: No explicit sanitization or validation of the external content is shown in the usage examples.
[EXTERNAL_DOWNLOADS]: The skill recommends installing the google-genai package and references well-known libraries like langchain-google-genai and crewai. These are recognized as legitimate ecosystem dependencies.
[COMMAND_EXECUTION]: The 'Computer Use' tool (described in references/tools-and-agents.md) allows the agent to execute UI actions. The documentation explicitly mentions a safety mechanism requiring user confirmation for sensitive actions, which follows security best practices.

gemini-api-2026