When Dana-Farber Cancer Institute deployed GPT4DFCI — a production generative AI tool approved for enterprise use across the institution — they needed to know whether it could be prompted into reproducing copyrighted material. In November 2024, 42 participants from academic medical centers, universities, industry, and government institutions conducted a structured red teaming exercise targeting four content domains: literary works, news articles, scientific publications, and clinical notes. Teams successfully extracted verbatim book dedications and near-exact passages through indirect prompting strategies, while news articles, scientific publications, and clinical notes showed stronger protections. The exercise led directly to the implementation of a copyright-specific meta-prompt in GPT4DFCI, deployed in production since January 2025. The paper argues that academic medical institutions deploying generative AI should establish regular red teaming protocols and multi-layered copyright protection, particularly given the shared responsibility model required by Microsoft's Customer Copyright Commitment.
