The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and act upon live data from external application user interfaces, creating a potential vector for indirect prompt injection.
Ingestion points: The skill retrieves application state and tool definitions using view-screen and list-host-webmcp-tools as described in SKILL.md.
Boundary markers: There are no explicit delimiters or instructions provided to the agent to treat data from the application screen as untrusted or to ignore embedded instructions within that data.
Capability inventory: The skill can execute powerful actions within the workspace apps using run-host-webmcp-tool and ask_app based on its interpretation of the screen state.
Sanitization: The instructions do not mention any validation, filtering, or sanitization of the content retrieved from the external application surfaces.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill configures a connection to an external network endpoint at https://dispatch.agent-native.com/mcp to facilitate its core functionality. While this domain is not on the standard whitelist, its use is transparently documented as the primary service endpoint. The skill mitigates risks by including strict prohibitions against requesting or storing sensitive user credentials such as passwords, OAuth codes, or session cookies.

an