The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a custom multimodal evaluation metric in reference/multimodal-eval.md that constructs prompts for a judge model by directly interpolating untrusted content. This pattern creates a surface where malicious instructions embedded in the data being evaluated could influence the evaluation results.
Ingestion points: actual.user_content.parts processed in multimodal_response_quality function.
Boundary markers: Not used; agent responses and user content are concatenated directly with judge instructions.
Capability inventory: The script uses a generative AI client (client.aio.models.generate_content) to process the resulting prompt.
Sanitization: No evidence of escaping or filtering of the interpolated content.
[DYNAMIC_EXECUTION]: The framework described in reference/criteria-guide.md supports the use of custom metrics by dynamically loading Python functions from string-based module paths defined in configuration files.
Evidence: The custom_metrics configuration uses code_config to specify module names such as my_app.eval.my_module.my_metric_function for runtime execution.
[EXTERNAL_DOWNLOADS]: The skill refers to external documentation and guidelines hosted on Google's GitHub Pages domain.
Evidence: Links to https://google.github.io/adk-docs/ are provided for further reference on evaluation criteria and user simulation.

adk-eval-guide