[2506.22523] Red Teaming for Generative AI, Report on a Copyright-Focused Exercise Completed in an Academic Medical Center

0 2 minutes read

250622523 Red Teaming for Generative AI Report on a Copyright Focused.png

[Submitted on 26 Jun 2025 (v1), last revised 2 Jul 2025 (this version, v3)]

Authors:James Wayne, Sahil Nalwadi, Chey Liang, Catherine Belek, Marissa Ferrara Boston, Alexander Chaudhry, Adele Collin, Luigi de Anjis, Jacob Ellen, Heather Farasi, Rodrigo R. Ann Cook, Yanan Los Lu, Heather Matti, Liam J. McCoy, Catherine Miller, Alison Morgan, Marilyn Louisa Mij, Tarng Ngwin, Alexander Owen Post, Alex D. Velez, Bengcheng Wang, Anna West, Yusang Chu, Yingdy Chu, Jason M. Johnson, Naomi Lennan, Jennifer Wilkox, Francis J.

View a PDF file from the paper entitled Red Teaming for Ai Touly, a report on an exercise that focuses on the copyright that has been completed in an academic medical center, by James Win and 40 other authors

PDF view

a summary:The background: AI’s spreading in academic medical environments raises compliance concerns about copyright. The Dana-Farber Institute for Cancer GPT4DFCI, the AI, is carried out by the internal obstetric tool that uses Openai models, which have been approved to use institutions in research and operations. Looking at (1) the exceptional adoption of the broad tool in our institution, (2) our research mission, and (3) the joint responsibility model required to benefit from the commitment of printing and publication rights in the Azure Openai service products, we considered the test of compliance with the necessary strict copyright.

Status Description: We have had an organized exercise for red groups in November 2024, with 42 participants from academic, industrial and governmental institutions. Four teams tried to extract protected content from GPT4DFCI through four areas: literary works, news articles, scientific publications, and clinical notes restricted to reach. The teams successfully extracted the dedication of literal books and clips near the excitement through various strategies. The news article has failed despite attempts to break the protection. The reproduction of the scientific material only resulted in high -level summaries. The clinical memo test revealed the appropriate privacy guarantees.

Discussion: A successful extraction of literary content indicates the presence of potential protected materials in training data, which requires the liquidation of the reasoning time. Differential success rates through the types of content indicate varying preventive mechanisms. This event has carried out the descriptive expression of copyright in GPT4DFCI; This mitigation has been under production since January 2025.

Conclusion: Systematic Red RED team revealed specific weaknesses in compliance with copyrights from artificial intelligence, which leads to concrete mitigation strategies. Academic medical institutions that spread artificial intelligence must implement the continuous testing protocols to ensure legal and moral compliance.