This research introduces D-SEC, a threat model that separates attackers from legitimate use and captures dynamic, multi-step interactions. Using a crowd- sourced red-teaming platform—we analyze 279k real-world attacks and show how some defenses degrade usability. We highlight effective strategies like adaptive defenses and defense-in-depth.