The Agent Skills Directory

[OBFUSCATION]: The file demo/index.html encapsulates the demo's HTML document and logic within a large Base64-encoded string (encodedHtml). This content is decoded at runtime using atob() and injected into a sandboxed iframe via the srcdoc attribute. This is a common architectural pattern for self-contained design previews. Analysis of the decoded content shows a standard design implementation including a Content Security Policy (CSP), animations, and WebGL code, with no evidence of malicious payloads.
[INDIRECT_PROMPT_INJECTION]: The skill instructions define a vulnerability surface by directing the agent to ingest and reference data from an external vendor-controlled URL.
Ingestion points: Instructions in demo/PROMPT.md direct the agent to fetch and analyze visual and interaction references from https://neuform.ai/pages/alert-interface/index.html.
Boundary markers: The prompt instructions do not include explicit boundary markers or directives to ignore instructions that may be embedded within the external reference content.
Capability inventory: The skill includes JavaScript for 2D canvas manipulation and procedural animation in SKILL.md, and the demo manages iframe content and asset injection via postMessage and srcdoc in demo/index.html.
Sanitization: There is no evidence of input validation or sanitization of the external data before it is processed by the agent.

mesh-gradient-dark-blue-clean