Staff-context script execution via unescaped LLM output

Stored XSS in affected Discourse 2026.1, 2026.2, and 2026.3 releases allows remote authenticated users to run scripts in staff browsers. The llm_triage and FlagPost paths insert raw LLM response and automation-name values into review flag reasons that the interface treats as HTML-safe, without first escaping them. Exploitation requires AI triage to process attacker-influenced post content and an administrator or moderator to open the resulting flagged post; the payload then executes in that staff member's browser context.