The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill renders markdown, math, and diagrams derived from assistant messages, which may include content influenced by untrusted external data.
Ingestion points: Untrusted data enters the agent context through the text part of message primitives (e.g., part.text in SKILL.md).
Capability inventory: The skill uses innerHTML to render Mermaid diagrams and utilizes rehype-katex for math rendering as described in references/latex-mermaid.md.
Sanitization: The StreamdownTextPrimitive documented in references/streamdown.md includes extensive security configurations: allowedTags for HTML whitelisting, security for URL/image prefix restrictions, and linkSafety for external link confirmation.
Boundary markers: The skill uses standard markdown syntax to delimit blocks but does not add explicit boundary markers to ignore instructions within the processed text.

markdown