The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external data including JSON manifests, SVG vectors, and video files, which presents a surface for indirect prompt injection if the agent processes untrusted asset content.
Ingestion points: External resources are loaded through Assets.load(), Assets.loadBundle(), and Assets.init() with remote manifest URLs, as seen in SKILL.md and references/manifests.md.
Boundary markers: The instructions do not define specific boundary markers for the data content within loaded assets.
Capability inventory: The skill enables network access to retrieve assets and performs parsing of complex formats like SVG (detailed in references/svg.md).
Sanitization: No specific sanitization procedures for external asset content are mentioned in the documentation.
[SAFE]: The code patterns and instructions are consistent with the intended functionality of the PixiJS v8 library. Example URLs use placeholder domains, and the base64 content in references/gif.md is a standard 1x1 GIF data URI. No malicious obfuscation, unauthorized data exfiltration, or privilege escalation patterns were detected.

pixijs-assets