Lockdown Mode and sandboxing reduce risk but every rendering feature added to an AI client is new attack surface. For teams: be wary of shared chats and custom GPTs from outside the org, keep sensitive data out of assistants that render untrusted HTML, and when you build agent UIs, sandbox previews without allow-popups / allow-popups-to-escape-sandbox and block window.open at the iframe level, not by wrapping the function.