The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides examples of agent architectures that interpolate user-provided or state-derived data directly into system and user prompts without explicit boundary markers. This is a standard operational pattern for LLM agents but represents an attack surface for indirect prompt injection.\n
Ingestion points: User messages and shared state (e.g., proverbs list) are ingested through the AG-UI protocol and processed by the agent (e.g., in references/integrations/adk.md and references/integrations/strands.md).\n
Boundary markers: The illustrative code snippets, such as before_model_modifier in adk.md and build_proverbs_prompt in strands.md, do not utilize delimiters or specific instructions to isolate or ignore potentially malicious instructions embedded within the interpolated data.\n
Capability inventory: The agents described in the guides have access to various capabilities including network requests (via weather and financial tools) and state modification (via update tools) (e.g., in references/integrations/adk.md, references/integrations/agno.md, and references/integrations/strands.md).\n
Sanitization: No explicit input validation, escaping, or sanitization of the interpolated state data is present in the provided documentation examples.

copilotkit-integrations