Alignment Whack-a-Mole: Finetuning Bypasses Safety to Unlock 85-90% Verbatim Recall of Copyrighted Books in LLMs

arXiv / GitHub / Hacker News·high signal

Researchers demonstrated that finetuning GPT-4o, Gemini-2.5-Pro, and DeepSeek-V3.1 on plot-summary-to-text tasks caused models to reproduce 85-90% of held-out copyrighted books, with single verbatim spans exceeding 460 words — using only semantic descriptions as prompts and no actual book text. Training exclusively on Haruki Murakami novels unlocked recall of books from 30+ unrelated authors, showing the bypass generalizes. Paper directly challenges RLHF safety alignment defenses. 174 points, 138 comments on HN.