The Agent Skills Directory

Indirect Prompt Injection Surface: The SDK processes untrusted input from users and external models, which is a standard characteristic of AI agent frameworks that can be leveraged for indirect prompt injection. The skill provides structure to help mitigate these risks, but they remain an area for developer review.
Ingestion points: External data is ingested through the messages list provided to the agent or model stream in SKILL.md.
Boundary markers: The SDK utilizes typed messages (e.g., ai.system_message, ai.user_message) to define the role and context of different inputs.
Capability inventory: The SDK supports a variety of operations including model streaming, tool execution, and multi-agent coordination.
Sanitization: As detailed in references/custom-loops.md, the context.resolve function handles argument validation and implements approval gates for tool execution.

ai