analiza aiqbench

This write-up details a chained web-security flaw in ChatGPT's HTML preview feature, where sandbox escapes via a third-party library and a permissive CSP let a crafted share link exfiltrate other signed-in users' data. The core lesson is that AI product features inherit classic web-security risks, so sandboxing, message validation, and CSP allowlists need the same rigor as any app handling private data. The $5,000 bounty is a useful reference point for how AI vendors currently price serious vulnerabilities.