The Agent Skills Directory

[OBFUSCATION]: The file demo/index.html contains a large Base64-encoded string assigned to encodedHtml which decodes into a complete HTML document containing embedded styling and script placeholders.
[DYNAMIC_EXECUTION]: The script in demo/index.html uses atob() to decode the embedded document and dynamically injects it into an iframe using the srcdoc attribute, which prevents direct static analysis of the rendered content.
[DYNAMIC_EXECUTION]: In demo/index.html, the runtimeFiles array contains several relative paths using path traversal sequences (../../../../) to access shared runtime libraries located outside the skill's local directory.
[EXTERNAL_DOWNLOADS]: The skill's demo configuration in demo/source.json and the decoded content of demo/index.html reference external assets and scripts from well-known technology services, including Google Fonts, Cloudflare (hosting GSAP, Three.js, and ScrollTrigger), Tailwind CSS, Iconify, and Supabase.
[INDIRECT_PROMPT_INJECTION]: The design instructions in demo/PROMPT.md encourage the agent to ingest content from an external URL (https://neuform.ai/pages/neus-enterprise-brand-architecture/index.html) to use as a high-fidelity visual reference.
Ingestion points: The agent is directed to use the external Neuform design page and the obfuscated demo HTML as visual and interaction references.
Boundary markers: No explicit delimiters or instructions are provided to the agent to treat the reference content as untrusted data.
Capability inventory: The skill environment utilizes an iframe with sandbox="allow-scripts" for dynamic execution of the demo content.
Sanitization: Content Security Policies (CSP) are implemented in both the wrapper demo/index.html and the obfuscated HTML payload to restrict resource loading.

editorial-tech