The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The framework described enables agents to process untrusted data from user inputs and external tool responses, creating an inherent surface for indirect injection.
Ingestion points: User content processed through Runner.run_async and data retrieved from external sources like google_search or MCPToolset.
Boundary markers: The documentation recommends using delimiters (e.g., XML tags or triple dashes) to isolate data from instructions, though enforcement is left to the developer.
Capability inventory: The framework supports a wide range of capabilities including file operations via ArtifactService, network requests, and database access through various toolsets.
Sanitization: Guidance is provided on implementing guardrail agents and lifecycle callbacks (before_model_callback) to filter and validate content.
[DYNAMIC_EXECUTION]: A technical reference for building custom tools includes an example that uses the Python eval() function to process mathematical expressions.
Evidence: return str(eval(expression)) in references/tools-reference.md.
Context: This is presented as an illustrative example for a calculator tool within a development guide.

adk-skill