The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides integration templates that create a surface for indirect prompt injection. Specifically, several examples demonstrate how to interpolate shared state (such as a list of proverbs) directly into the agent's prompt without using boundary markers, delimiters, or sanitization logic.
Ingestion points: Shared state is ingested via callback_context.state (references/integrations/adk.md), input_data.state (references/integrations/strands.md), and initial_state (references/integrations/llamaindex.md).
Boundary markers: Code patterns such as the build_proverbs_prompt function in strands.md and the before_model_modifier in adk.md append state data directly to the instructions, which could allow malicious content within that state to influence agent behavior.
Capability inventory: The agents described in the examples have access to various tools (e.g., set_proverbs, get_weather, financial tools) and can interact with frontend components.
Sanitization: No input validation or escaping of the shared state data is included in the provided implementation snippets.

copilotkit-integrations