The Agent Skills Directory

[EXTERNAL_DOWNLOADS] (LOW): The skill documentation instructs the user to install several NPM packages from the @streamdown scope. While these are not from the pre-approved trusted list, they are essential to the skill's primary purpose. Severity is downgraded from MEDIUM to LOW as per the use-case rule.
[DATA_EXFILTRATION] (SAFE): No hardcoded credentials or sensitive file access patterns were found. The skill actively mitigates exfiltration risks through a 'Link Safety' confirmation modal and rehype-harden for protocol/domain restrictions.
[REMOTE_CODE_EXECUTION] (SAFE): No patterns of piping remote content to a shell (e.g., curl|bash) or dynamic execution of system commands were detected. Diagram rendering (Mermaid) and Math (KaTeX) are executed client-side.
[INDIRECT_PROMPT_INJECTION] (LOW): As a renderer for AI-generated text, the skill represents a surface for indirect prompt injection. A malicious AI response could attempt to exploit the rendering engine or use social engineering via the rendered output.
Ingestion points: Untrusted data enters via the children prop of the Streamdown component (e.g., assets/examples/basic-streaming.tsx).
Boundary markers: None explicitly defined in the provided component usage, though standard React prop boundaries apply.
Capability inventory: The skill allows rendering of raw HTML (via rehype-raw), Mermaid diagrams, and KaTeX math.
Sanitization: Robust sanitization is provided via rehype-sanitize and rehype-harden by default.

streamdown