The Agent Skills Directory

[DATA_EXFILTRATION]: Indirect Prompt Injection Surface. \n
Ingestion points: The skill utilizes makeAssistantVisible to ingest rendered HTML (references/visible.md) and useAssistantContext to ingest application state strings (references/model-context.md). \n
Boundary markers: Documentation does not mention the use of delimiters or instructions to ignore embedded commands in ingested content. \n
Capability inventory: The framework provides tools for the agent to interact with the UI, such as click and edit tools (references/visible.md), and state mutation tools like update_{name} (references/interactables.md). \n
Sanitization: No sanitization or validation of the content read from the DOM or application state is described before it is provided to the agent.\n- [DATA_EXFILTRATION]: The makeAssistantVisible utility exposes the outerHTML of React components to the AI model's context. This may lead to the exposure of PII or other sensitive data if such information is rendered in the component tree (references/visible.md).\n- [COMMAND_EXECUTION]: The skill documentation describes tools that allow the agent to interact with the UI by clicking elements and editing input fields, as well as mutating application state through auto-generated update tools (references/interactables.md).\n- [EXTERNAL_DOWNLOADS]: The skill references and directs users to official documentation and API resources at assistant-ui.com, which is the vendor's primary domain (SKILL.md).

copilots