The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external topology data (hostnames, roles, and link relationships) to generate a text prompt for a 3D world generator. This creates a surface for indirect prompt injection where malicious topology data could attempt to influence the generative model.
Ingestion points: Topology data is ingested into the TopologySnapshot object in SKILL.md workflows.
Boundary markers: The fantastical_prompt_builder.build_prompt function encloses the dynamic content within a fixed text template to guide the model.
Capability inventory: The skill utilizes the worldlabs-marble-mcp/generate_world tool to send data to an external API.
Sanitization: topology_model.py includes a sanitize_metadata function to strip credentials and configuration blobs, and the prompt builder specifically limits input to hostnames and roles to avoid accidental data leakage.
[DATA_EXFILTRATION]: The skill transmits network hostnames and connectivity patterns to the World Labs Marble API (platform.worldlabs.ai) to generate visualizations. While this is the primary function of the skill and the destination is a well-known service, it involves sending internal network identifiers to a third-party domain outside the standard whitelist.

worldlabs-topology-viz