Security Model (HTML + CSP)

HTML is an XSS surface

If you enable raw/inline HTML (features.html), treat the input as untrusted by default:

Keep sanitization enabled.
Prefer hosted workers (parsing off the main thread; clearer CSP posture).
Only extend allowlists/schemas for content you trust.

CSP recommendations

Prefer hosted workers

Host the worker bundle and point StreamMDX at it:

<StreamingMarkdown worker="/workers/markdown-worker.js" />

This avoids needing to allow blob: in worker-src/script-src.

Avoid `unsafe-eval`

StreamMDX does not require unsafe-eval for its core pipeline. If your MDX components embed third-party scripts, treat those as a separate surface and lock them down per-component.

Reporting security issues