The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines patterns for fetching data from external APIs and reading scene files, which introduces an attack surface for indirect prompt injection if the ingested data contains instructions intended to influence the agent.
Ingestion points: fetch, signedFetch, and readFile functions described in SKILL.md.
Boundary markers: The examples do not specify the use of delimiters or warnings to ignore instructions embedded within the ingested content.
Capability inventory: The skill enables high-privilege actions including network communication (fetch, WebSocket), browser redirection (openExternalUrl), world navigation (teleportTo), and spawning external experiences (spawn).
Sanitization: Usage examples show standard data parsing (e.g., JSON.parse) without specific LLM-focused sanitization of the content.
[EXTERNAL_DOWNLOADS]: The skill provides references and links to example scenes and testing resources hosted on Decentraland's official GitHub organization.

scene-runtime