The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides detailed instructions and code examples for using shader modifiers and SCNProgram. These allow the injection and runtime compilation of GLSL and Metal Shading Language (MSL) code into the rendering pipeline.
Evidence: references/scenekit-patterns.md demonstrates assigning GLSL strings to material.shaderModifiers and creating SCNProgram objects with custom Metal function names.
[DATA_EXFILTRATION]: The skill includes multiple links to documentation hosted on a non-whitelisted domain (sosumi.ai). While these appear to be documentation links, they target a non-standard external service.
Evidence: SKILL.md contains links such as https://sosumi.ai/documentation/scenekit and https://sosumi.ai/videos/play/wwdc2025/288/.
[INDIRECT_PROMPT_INJECTION]: The skill describes methods for loading external 3D models (.usdz, .scn, .obj) and serialized scene data. This creates a vulnerability surface where malicious instructions could be embedded in asset metadata or properties, especially given the skill's capability to execute dynamic shaders.
Ingestion points: SKILL.md (Loading Models section) and references/scenekit-patterns.md (Scene Serialization section).
Boundary markers: No delimiters or instructions to ignore embedded content are provided.
Capability inventory: The skill can execute arbitrary shader code (via shaderModifiers) and perform file system operations (via scene.write(to:...)).
Sanitization: No evidence of validation or sanitization for external assets is present.

scenekit