• Sources: The Hacker News report
  • Summary: Parameter reported the flaw on 2026-10-18 and OpenAI fixed it by 2026-10-26, rating it P1. A shared chat with an HTML Preview and an interactive graph let attacker HTML reach the main page, build an unsandboxed same-origin frame and load code through a YouTube script that the content security policy allowed. No real-world use was reported.
  • Why it matters: Products that render model-generated HTML must check postMessage senders in every library loaded on the trusted page, not only harden the sandbox.