The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external 3D model files and references documentation from an unofficial domain, creating a surface for malicious instructions to be ingested by the agent.
Ingestion points: The skill loads scene data from external files via SCNScene(url: ...) in references/core-implementation.md and references/scenekit-patterns.md. It also references documentation at https://sosumi.ai/documentation/scenekit.
Boundary markers: The skill suggests using .checkConsistency: true and a loadCheckedScene function to verify node presence, though these are structural checks rather than security filters.
Capability inventory: The skill has the ability to read 3D scene files, render content via SCNView, and write scenes to the file system using scene.write(to: ...).
Sanitization: Includes basic consistency checks during the scene loading process.
[DATA_EXPOSURE]: The skill interacts with the local file system and references non-standard external domains for documentation.
Evidence: The skill demonstrates writing scene data to the temporary directory using FileManager.default.temporaryDirectory.appendingPathComponent("scene.scn") in references/scenekit-patterns.md.
Evidence: References to the sosumi.ai domain (e.g., https://sosumi.ai/documentation/scenekit) are included, which is not an official Apple documentation source and represents an unverified network destination.
[DYNAMIC_EXECUTION]: The skill uses runtime code injection for 3D rendering effects.
Evidence: Uses SCNProgram to execute custom Metal shader functions and shaderModifiers to inject GLSL-like snippets into the rendering pipeline as shown in references/scenekit-patterns.md.

scenekit