AI-generated content is now sanitized before it renders

Requirements documents, lead descriptions and proposal summaries can all be drafted by an AI agent, and that content is rendered as rich text rather than plain text. Until now the step that was meant to make it safe to render did not actually do so — it tidied up the formatting an AI model sometimes wraps its answer in, and passed everything else through untouched.

That step now applies a real allowlist: only the tags rich text genuinely needs survive, and anything that could run code in the browser is removed. It covers every place AI-drafted content is displayed, so a document does not become a risk depending on which screen it is opened from.

Alongside it, the agentic assistant now keeps its own instructions and behaviour in its own application, and the assistant app can be restricted to a named list of sign-in origins. Neither changes what you see today; both make the agent work now under way safer to build on.

One known effect

The new allowlist covers the standard set of rich-text formatting. Content saved with formatting outside that set — a top-level heading, an underline, an inline image or a code block — will show without that formatting when you view it, while still appearing correctly if you reopen it in the editor. Nothing you have written has been changed or lost. Widening the allowlist to match everything the editor offers is already queued as the next fix.