Parameter researchers disclosed a flaw allowing a malicious shared ChatGPT conversation to execute code in a signed-in viewer’s session and read private conversation lists and messages. The chain combined unauthenticated cross-frame messages in MathLive, HTML insertion and an unsandboxed frame; an automatic-preview variant required no further click. According to the researchers, OpenAI confirmed fixes in both affected web frontends on October 26. This was a research demonstration of data access, not a report of widespread real-world theft. The defensive lesson is to verify both message origin and sender window and preserve sandboxing, rather than relying on content security policy alone.